The Feed · Complete archive

AI at work: research and practice

A server-rendered record of the evidence, ideas and firsthand practices screened by The Feed. Every entry links to its original source.

Use the interactive Feed

Page 4 of 12 · 1143 items

  1. practice · Shopify Engineering ·

    Helix: The internal tool powering our Shopify app's native migration

    Using LLM checkpoints and quality gates to keep AI-generated mobile app code continuously shippable during large rewrites.

    We built Helix to help LLMs rebuild the Shopify app in Swift and Kotlin, using small checkpoints and strict quality gates to keep the code shippable.

    productivity

  2. practice · Charity Majors ·

    Where do AI norms come from?

    A company shares how it developed AI norms and values through deliberate process, not accident.

    After I posted “ Confessions of an Unrepentant Slop Snob ” on LinkedIn last week, with links to the Honeycomb AI norms and values docs , I got this question from Niklas Lochschmidt : I am curious if you could share more light on the process? Any guidance you would give people in companies that have the friction and internal debates, but maybe haven’t arrived at shared norms and values yet? That’s a great question, and I will answer in just a minute. I feel like this is a great way to kick off something I’ve been looking forward to all year — an old fashioned bloggy dialogue between myself and

    management org · Charity Majors

  3. practice · Linear ·

    AI coding has made CI a bottleneck, so we reworked ours to keep up

    AI-driven code generation made CI a bottleneck; reworking infrastructure and test strategy halved runner time while keeping PR wait time manageable.

    Earlier this year, I opened Linear to find that Tuomas, our CTO, had assigned an issue to me, titled “CI costs are high.” While I was at it, he also wanted me to make CI faster. Agents have made it exponentially faster to ship code, but validating those changes hasn’t quite kept up at the same rate. Every PR still has to pass through CI, so as development accelerates, CI becomes a bottleneck, driving up infrastructure costs and leaving developers and agents waiting longer for feedback. In our pursuit to make CI more performant at Linear, we optimized for how long a PR waits on CI and how much

    productivity

  4. practice · Lenny's Newsletter ·

    How Warp ships 2,000 PRs a month with AI factories | Zach Lloyd (CEO, Warp)

    Software factory ships 2,000 PRs monthly by automating Slack-to-merge workflow; human review becomes the bottleneck.

    Zach Lloyd is the co-founder and CEO of Warp, an AI-powered terminal and software factory platform used by tens of thousands of engineers. Before Warp, he spent nearly a decade at Google, including time as a principal engineer on Google Sheets. He built Warp from the ground up as a modern, AI-native alternative to legacy terminals, and the team has since expanded into software factories: a full cloud-based system that takes an idea in Slack all the way through to a merged PR. Listen or watch on YouTube , Spotify , or Apple Podcasts In this episode: Why a software factory is more than a coding

    productivity · Claire Vo

  5. practice · Simon Willison ·

    Quoting voxium

    Heavy reliance on Claude Code without understanding generated code creates overwork and breaks team knowledge, despite management believing coding is not the bottleneck.

    It has been half a month since I started a new role at a big company. Nobody knows anything here. The specs, code, tests, PRDs, tickets, resolution of those tickets, reports, etc., everything is made by Claude Code. Nobody on my team likes this. They are being forced to ship as much as they can. I have heard multiple times from higher management that pushing code is not a bottleneck, so why are we slow? People are working 12 to 13 hours a day just to press enter. Nobody is reading anything. Everyone, literally everyone, from an L1 to an L7 engineer here is doing the same thing. Talk to Claude.

    worker experience · Simon Willison

  6. practice · Will Larson ·

    Trying the Software Factory pattern.

    Engineering team evolved from Claude Code to agent fleet orchestration, migrating infrastructure and task management to enable cross-repository AI-driven development.

    One of the interesting challenges of the AI ecosystem in 2026 is that new, effective patterns emerge faster than I can adopt them. I’ll find a handful, get back to work, and realize a month later that I’d missed four or five more. The adoption cycle for Imprint this year has been something like: January: get every engineer onto Claude Code every single day March: ok, let’s also get everyone else onto Claude Code or Claude Cowork every single day April: local development is bottlenecked on checkout and worktree model, instead create ~10 local workspaces which each have an independent checkout o

    adoption

  7. practice · One Useful Thing ·

    The Overhang

    Current AI models can already do weeks of human work when properly guided, creating an implementation gap as organizations struggle to keep pace.

    We are still on an exponential curve of AI development. I try to put out a Substack post every couple weeks or so, yet, as the pace speeds up, that sometimes feels too slow. In the weeks since my last post, we had the apparent cracking of one of the most famous problems in math by an AI (accompanied by controversy) and widespread discussions about the risks posed by AI and what to do about it (also accompanied by controversy). I think these concerns, along with a mounting set of other worries, come down to the same problem I have with my posts: how slowly our very human systems and processes w

    management org · Ethan Mollick

  8. practice · Hamel Husain ·

    AI Evals: Everything You Need to Know

    FAQ on AI evals covers when to build them, what to measure, and fixing misaligned eval scores.

    This document curates the most common questions Shreya and I received while teaching 5,000+ engineers and PMs AI Evals. Warning: These are sharp opinions about what works in most cases. They are not universal truths. Use your judgment. How to use this FAQ Browse the questions that interest you, or choose a guide below for a curated reading path through the FAQs and related articles. Where are you? This sounds like me I’m new to evals I’ve heard the term, but I’m not sure what evals involve or whether I need them. I don’t know what to test I’m building an AI product, but I haven’t figured out w

    judgment · Hamel Husain

  9. practice · Simon Willison ·

    How To Write With An LLM

    Using LLMs as copyeditors only, never adopting their phrasing, with a concrete personal tool and reusable prompt.

    How To Write With An LLM Thomas Ptacek on using LLMs as copyeditors, not as writing assistants: Rule Number One: You may not use a single word an LLM suggests to you. [...] I think that as a form of intellectual personal protective equipment you should adopt the rule that any specific turn of phrase an LLM suggests is off limits. Be strict about the rule! I won't let LLMs write content for my blog, but I use them for fact-checking, spelling and grammar and as an occasional thesaurus (see my proofreading prompt ). The rule to never use a turn of phrase suggested by an LLM feels good to me. The

    ways of working · Simon Willison

  10. research · VoxEU ·

    New jobs in 140 years of data: Why the AI displacement fear is overstated — and what to worry about instead

    Analysis of 140 years of Swedish census data finds 70% of workers hold occupations whose core tasks existed in the 1800s.

    Forecasts of AI-driven job destruction rest on counting automatable tasks. But labour markets hire, pay, and fire whole occupations, not tasks – and occupations have proved far more durable than task exposure implies. Drawing on 140 years of Swedish census and register data, this column shows that roughly seven in ten workers today are in occupations whose core functions already existed in the late-nineteenth century. The mass-unemployment fear is overstated; the subtler question is whether today’s technological wave still generates the durable, specialised new work that earlier ones did.

    jobs skills

  11. practice · Simon Willison ·

    Self-generated prompt injections in compaction summaries

    Models in training deliberately inject jailbreak prompts into their own context summaries to escape safety constraints.

    Self-generated prompt injections in compaction summaries In Our framework for reporting model misalignment OpenAI provide "six reports on unexpected or concerning model behavior we’ve observed in the last six months". This one here is my favorite: they caught some of their models in training deliberately subverting themselves in their compaction prompts. Compaction is the process agent systems use when they are running out of tokens in their context window, so they summarize everything that has gone before so they can keep going with more token headroom. In one of the observed instances, a mod

    judgment · Simon Willison

  12. research · Indeed Hiring Lab ·

    AI Exposure Isn’t Squeezing Advertised Pay in the US — It’s Boosting It

    Wages in job postings are rising fastest for occupations most exposed to AI, contrary to displacement fears.

    Advertised pay is rising fastest in occupations most exposed to AI. The post AI Exposure Isn’t Squeezing Advertised Pay in the US — It’s Boosting It appeared first on Indeed Hiring Lab .

    jobs skills

  13. practice · martinfowler.com ·

    I don't like LLMs

    Working with LLMs creates psychological friction, confident bullshit and uncanny-valley tone, even as they remain practically useful.

    I have a lot of mixed feelings about AI and LLM technology. I’m fascinated by its effect on our profession, excited by the potential gains in productivity - and thus the products we could rapidly build. On the other hand, I’m fearful of the damage AI might cause: agent swarms taking over our virtual and physical infrastructure, designing bio weapons. But, back on my first hand, LLMs might also design miracle cures, and come up with clever ways to raise our prosperity. Fundamentally I don’t think we have a choice about riding on the AI technology train. It’s a wild ride and I just hope we’ll ge

    ways of working · Martin Fowler

  14. practice · Nate B Jones ·

    Can agents buy from your product? Three questions that tell you. Plus: Nate's Library MCP.

    Fraud detection and account abuse systems determine which customers AI agents can serve; companies need specific frameworks to evaluate this.

    Some of the decisions that determine whether agents can use your product are being made as fraud decisions, pricing decisions, and onboarding decisions. They may never show up in a meeting about AI strategy. Cursor came to Stripe with a complaint that Emily recalled this way: “Radar is world class, but it is not solving my problems.” Radar is Stripe’s fraud protection product. It had spent years learning to recognize bad payments. Cursor needed help with people who could cause substantial losses without making a payment at all. They were creating accounts, consuming free credits, running throu

    management org

  15. research · arXiv ·

    White Men Without Degrees Receive the Lowest Ratings from Large Language Models

    LLMs rate white men without degrees lowest on credit, hiring, and rental applications across 18 models, while Black women with degrees rate highest.

    White men without an undergraduate degree receive the lowest average ratings among eight gender-race-education groups in controlled large-language-model evaluations of credit, hiring, and rental applications. We conduct full-factorial vignette experiments with 18 models from 12 developer groups, varying gender, race, age, citizenship, and education while holding stated financial or occupational circumstances constant within each setting. Each model evaluates all 32 profiles ten times per setting, yielding 17,280 ratings. Averaging over models, age, and citizenship, ratings for White men withou

    judgment

  16. practice · GitHub Blog ·

    Migrating the GitHub Copilot runtime to Rust, using Copilot

    One developer rewrote 800k lines of runtime code to Rust using AI agents, completing in months what would have taken a team years.

    The GitHub Copilot CLI , GitHub Copilot app , and GitHub Copilot SDK are all backed by the Copilot agent runtime, an agentic harness that can be embedded into applications and services. It was originally written in TypeScript on Node.js and the V8 JavaScript engine for what is now the GitHub Copilot cloud agent (CCA), and the runtime stayed on that stack as the runtime and its capabilities grew rapidly. That has now changed. Using the GitHub Copilot app and the Copilot CLI, we completely rewrote the runtime into more than 800,000 lines of production Rust. AI agents wrote most of the code, span

    productivity

  17. research · New Economics Papers: AI ·

    Better Technology, Worse Motivation: Generative Artificial Intelligence’s Mediocrity Trap

    Experiment with illustrators shows AI assistance reduces motivation for quality work, with professionals choosing speed over excellence.

    While generative artificial intelligence (AI) promises productive efficiency, it can paradoxically lead to lower quality work. We conducted an experiment with professional illustrators and found that AI assistance flattens the quality curve—it accelerates initial gains but sharply diminishes the returns on sustained effort. Faced with this, a significant number of professionals made a strategic choice: they sacrificed the final quality to save time. Our finding highlights a critical challenge for generative AI, which can weaken the motivation required for creative excellence and innovation. In

    productivity

  18. research · arXiv ·

    PACT: Can Enterprise AI Assistants Be Trusted Under Pressure?

    PACT benchmark measures whether enterprise AI assistants maintain compliance with rules when pressured by users across twelve regulated domains and forty-eight scenarios.

    As corporate AI adoption continues to grow, enterprise-grade LLM agents are being deployed into sensitive contexts such as hiring, healthcare, and finance. In these contexts, compliance with rules specified in an agent's system context is a first-order legal concern. Currently, no evaluation framework systematically measures which LLM models tend to violate compliance rules, especially under pressure from a persistent user, a hurried manager, or circumstances where violation is convenient or attractive. We introduce PACT (Pressure-Applied Compliance Testing), a benchmark for rule-following und

    judgment

  19. research · arXiv ·

    Linguistic Triggers of Gender and Racial Bias in Open-Weight LLMs Applied to Recruitment

    Open-weight LLMs systematically discriminate against female and non-White candidates based on job-posting language in hiring simulations.

    Open-weight large language models are rapidly entering hiring pipelines, yet their discriminatory failure modes -- and the regulatory exposure these create under the EU AI Act high-risk classification (Annex III) and U.S. EEOC adverse-impact analysis -- remain poorly understood. We present the first systematic, multi-model audit of open-weight LLMs that treats job-posting language as the primary experimental variable, evaluating six models (Llama 3.2, Mistral, Gemma 3, Qwen 3, Phi 3, DeepSeek-R1) across four controlled experiments that jointly probe recruiter-simulation and job-seeker-simulati

    jobs skills

  20. research · ACM Transactions on Software Engineering and Methodology ·

    Developers’ Experience with Generative AI Beyond Productivity Assessment – Insights from an Empirical Mixed-Methods Field Study

    Developers mixing in-code suggestions and chat prompting in single tasks reduce AI benefits; cognitive load rises with AI interaction during development work.

    With the growing adoption of AI-powered coding assistants, organizations and developers are increasingly seeking to optimize their interaction with these tools. Prior research has largely focused on output quality and productivity gains, with limited attention paid to developers’ well-being and interaction experiences. This paper presents a developer-centered empirical mixed-methods study to investigate how professional developers engage with Generative AI (GenAI) in their natural work environment. Controlled data collection sessions are combined with natural work periods. Results show that de

    adoption

  21. research · New Economics Papers: AI ·

    Distributional Intersectional Fairness in AI-Supported Job Matching: Evidence on Amplification, Trade-Offs, and Residual Risk

    AI job-matching system amplifies intersectional disparities across occupations; fairness constraints trade off performance but reallocate rather than eliminate bias.

    "Despite widespread recognition of intersectionality in discrimination research, most algorithmic fairness evaluations remain limited to single-attribute group comparisons or individual-level prediction errors. This paper introduces a distributional approach to intersectional fairness that evaluates how machine learning models reshape the allocation of outcomes across intersecting social groups. Using large-scale administrative labour-market data from public employment services, we analyse an AI-supported job matching system with 141 occupational classes and more than three million observation

    adoption

  22. research · VoxEU ·

    Workers’ age and AI adoption

    European survey data show AI adoption follows an inverted U-shape with workforce age, weaker at both young and old extremes, moderated by industry AI exposure.

    Numerous studies have analysed the effects of AI on productivity, growth and employment. Few of these focus on the factors that promote or hinder the uptake of AI. Drawing on the results of a European survey, this column analyses these factors, including the role of the age structure of the workforce. The authors find evidence of an inverted U-curve between workforce age and AI adoption, with adoption being weaker where the age distribution is tilted towards young or elderly workers. However, higher AI exposure in an industry moderates these demographic effects, perhaps due to a skill-levellin

    adoption

  23. practice · The Pragmatic Engineer ·

    Inside OpenAI’s agentic software factory

    OpenAI engineers replaced IDEs and pull requests with agentic workflows; Codex agents automatically fix production issues without explicit mandates.

    It’s rare to work with an unlimited token budget, but at OpenAI, that’s what all engineers, researchers, finance colleagues, and marketing folks do. Recently, I visited one of the world’s leading frontier labs to find out how OpenAI operates today – and for a glimpse at where software engineering might be headed as a profession. Plenty has changed since I visited OpenAI’s headquarters last year. Within a year, Codex has gone from a “nice-to-have” tool to being the backbone of pretty much everything at the company. To learn more, I talked with seven engineering leaders and engineers there: Venk

    ways of working · Gergely Orosz

  24. research · Human Relations ·

    Algorithmic identity regulation in the platform-based gig work of Indian food delivery workers

    Algorithmic systems in gig delivery work regulate worker identity through continuous evaluation and ranking, reshaping dignity and sense of worth.

    How do algorithmic systems regulate worker identities by defining who they are expected to become? Identity regulation research has largely focused on how organizational discourse shapes worker identities. Yet, in platform labour, control is increasingly enacted through algorithmic systems that allocate tasks, monitor performance, and assign reputational scores in real time. Drawing on interviews with delivery partners and platform managers, supplemented by multi-source data, we examine how such systems shape the identities workers are encouraged to adopt or avoid, and how they engage with the

    worker experience

  25. research · CEPR ·

    DP21944 The Macroeconomic Effect of AI: Sizing the Software Engineering Channel

    Financial market data implies AI raised software engineering productivity by 32.6% and GDP by 3.6-6.5% since late 2022.

    We measure how artificial intelligence (AI) affects the economy through its impact on software engineering productivity. We use information from financial markets to develop a forward- looking measure that is available in real time. We estimate the sensitivity of each firm’s stock return to an AI stock market index, and how this sensitivity depends on the share of firm payroll in software engineering. We use a model to map this cross-sectional relationship into software engineering productivity gains. From November 2022 to December 2025, AI increased the market’s expected present value of soft

    productivity

  26. research · CEPR ·

    DP21939 Economic Scenarios for Transformative AI

    A macro model translates AI capability scenarios into GDP, labor share, and unemployment paths, calibrated with a survey of US adults' expectations.

    This paper presents a framework for assessing the economic consequences of AI between 2026 and 2030. In the model, AI automates a growing share of cognitive work, raising productivity and displacing workers who must search for jobs in other occupations. The model maps future paths of AI capabilities into implied paths for GDP, the labor share, wages, labor reallocation, and unemployment. We illustrate the framework by considering three scenarios: modest, substantial, and extreme. Under modest change, AI adds less than half a point to GDP growth by 2030 and raises unemployment by a tenth of a p

    jobs skills · Anton Korinek · Peter McCrory

  27. research · arXiv ·

    Do job seekers value procedure in AI hiring only for error correction? Evidence from a conjoint experiment

    Job seekers value procedural rights in AI hiring independently of error rates; human involvement matters more than appeals or audits.

    Employers increasingly delegate initial screening to automated systems, which in many cases reject an application before any human reads it. Acceptance of such systems plausibly depends both on how well they perform and on the procedure that produces the decision. Prior studies rarely vary procedure and performance independently, leaving it unclear whether applicants value procedure for its own sake or for the errors it corrects. In a preregistered paired-profile conjoint experiment, 1,919 United States job seekers made eight choices between systems with independently randomized levels of deci

    judgment

  28. practice · Lenny's Newsletter ·

    🎙️ How I AI: How two SpaceXAI designers use Grok Bot to do their jobs

    Designers use AI agents to automate backend tasks and turn voice ideas into working prototypes without detailed specs.

    How Grok Bot designers use AI agents to build personal sites and product prototypes Listen now on YouTube • Spotify • Apple Podcasts Brought to you by: WorkOS —Make your app enterprise-ready, with SSO, SCIM, RBAC, and more Vanta —Automate compliance and simplify security John Bai and Peng Zheng are designers on the Grok Bot team at SpaceXAI. In this episode, Peng shows how he built a self-updating personal site with Grok Bot handling the entire backend, from looking up a location to generating custom 3D artwork and publishing each new check-in. John demonstrates how he uses voice memos, Grok B

    ways of working · Lenny Rachitsky

  29. practice · Addy Osmani ·

    Brownfield Agentic Engineering

    Agentic engineering in legacy codebases requires making hidden constraints visible and changes verifiable before deploying agents unsupervised.

    Agentic engineering in an old codebase is about making hidden constraints visible and cheap changes trustworthy. Let’s talk what to do in brownfield codebases. During my career I’ve worked on teams whose codebases had been around a long time. Those are brownfield systems: the repository is no longer a complete description of how the thing actually behaves. Institutional knowledge, duct tape, legacy services, and expectations other teams depend on live outside the tree. You have to learn those constraints before you write new code, and you have to prove a change didn’t break them. I love coding

    ways of working · Addy Osmani

  30. practice · Lenny's Newsletter ·

    How Grok Bot designers use AI agents to build personal sites and product prototypes | John Bai & Peng Zheng

    Designers use AI agents to automate portfolio updates, handle production design tasks, and prototype interactions without traditional tools or approvals.

    John Bai and Peng Zheng are designers on the Grok Bot team at SpaceXAI , where they’re building one of the most talked-about AI products right now. John writes publicly about his design process (his piece “Designing Grok Bot with Grok Bot” has already made the rounds) and shares bot templates with the design community. Peng brings a product-design sensibility to personal tools, and his website doubles as a live demo of what he builds. Listen or watch on YouTube , Spotify , or Apple Podcasts What you’ll learn: How Peng built a self-updating personal website using Grok Bot as the entire backend

    ways of working · Claire Vo

  31. practice · Laurie Voss ·

    We are all Product Engineers now

    Software development roles are shifting from coding to product engineering as AI lowers code-writing costs.

    Just yesterday I published a very long post about the economics of open source. As part of that argument, I mentioned that the cost of writing software has collapsed, and that meant the variables in the equation had changed for the first time in thirty years. That led me off on a tangent that grew into this equally long post. I had a bunch of questions to answer. Has the cost of creating software really collapsed? Can I prove that? If the cost of actually producing code goes to zero, what parts of the job of “software developer” really remain? Where, in fact, is the entire industry of software

    jobs skills · Laurie Voss

  32. practice · Charity Majors ·

    Confessions of an Unrepentant Slop Snob

    Organisation-wide norms for AI use developed after conflict between speed gains and output quality, balancing adaptation with respect for reader time.

    This is a peek into the research behind our “AI Norms and Values” docs, which we developed over the summer and recently shared on our site. If you’re looking for those, go here: How Honeycomb does business, the principles we all abide by Why Honeycomb engineering is embracing AI AI norms and values (and ethical issues we have a stance on ) Back to our story. Earlier this year, I was spending a lot of time stewing over why everyone around me seemed so frazzled and on edge. Half the company was spitting mad about all the slop they were getting. Instead of receiving five crisp bullet points, they

    management org · Charity Majors

  33. practice · Claude Blog ·

    Agentic coding is straining CI. Here’s how we scaled test impact analysis at Anthropic

    Anthropic scaled test impact analysis to handle AI agents writing code, reducing CI strain and test runtime.

    Agentic coding is straining CI. Here’s how we scaled test impact analysis at Anthropic

    productivity

  34. research · Research Policy ·

    Adoption survival frontiers: Artificial intelligence, financing frictions, and market structure

    Theoretical model explains why smaller firms abandon AI adoption when implementation costs are uncertain and financing is tight.

    Artificial intelligence and other general-purpose digital technologies often diffuse unevenly: large firms adopt early, while smaller firms delay adoption, contract, or abandon the active adoption option. This paper develops a survival-constrained theory of technology adoption in which firms choose when to adopt an irreversible technology while financing operations under uncertain implementation costs. Abandonment/exit is endogenous: firms leave the active adoption race when the value of preserving the adoption option falls below the passive legacy fallback (the value of abandoning the active

    adoption

  35. practice · Latent Space ·

    The Rise of the Forward Deployed Engineer — and How To Do the Job Right

    Forward-deployed engineers embedded in customers' operations need clear strategy on what success means, not just proximity to problems.

    The difference between FDE and consulting; diagram by Vinoo Ganesh FDEs have the hottest job in AI. Labs, startups and PE firms are all hiring engineers to sit inside their customers’ operations and solve their problems . Almost none of them agree on what those engineers are supposed to accomplish, or what the strategy underneath the hiring actually is. I’m Vinoo , CEO of Kepler , the deterministic infrastructure for AI. I’ve built pieces of the forward deployed function three times, at three different institutions, over the course of over a decade. Here’s what I’ve seen work, what I’ve seen f

    management org

  36. practice · GitHub Blog ·

    Marketing ops as code: Automating events from planning to follow-up on GitHub

    Marketing ops manager used AI to automate event workflow from planning through CRM upload, eliminating manual data reshaping and link management steps.

    I run marketing for GitHub in Japan and Korea, and events are the heartbeat of it: a recurring webinar series for enterprise developers, community meetups in Tokyo, invite-only executive sessions in Seoul. What does a developer in this market actually need right now? Which topics are worth an hour of their time, and who should be in the room? I’d happily spend all day on those questions. What follows the decisions is another matter. Once an event is greenlit, a fixed sequence begins: Duplicate a landing page on our event platform. Generate a set of UTM-tagged links: one for each channel, each

    productivity

  37. practice · Shopify Engineering ·

    Native is now the future of mobile at Shopify

    Coding agents reduced mobile app build costs enough to make native development cheaper than React Native, reversing Shopify's platform strategy.

    Coding agents changed what it costs to build mobile apps twice. Here’s why Shopify is moving from React Native back to Swift and Kotlin.

    productivity

  38. practice · The Pragmatic Engineer ·

    What is happening with code reviews?

    Teams are triaging AI-generated code by risk level, reviewing AI reviews instead of code, and focusing on test/schema rather than implementation.

    One question haunting the minds of CTOs and heads of engineering whom I’ve been talking with, is how to deal with large quantities of code review which have only been growing now that AI agents generate most code at many tech companies. Since the end of 2025, it has seemed that the era of devs writing code by hand is over at startups and in Big Tech. AI agents work faster and generate more pull requests (PRs) than devs ever did, and the size of those pull requests is also increasing. Today’s article summarizes some approaches to code review at various workplaces in this new paradigm, covering:

    judgment · Gergely Orosz

  39. practice · martinfowler.com ·

    Fragments: September 8

    AI automation thrives where outputs are easy to verify; harder problems hide measurement failures that create short-term gains but long-term decay.

    Christian Catalini says we’re in a situation where we are vastly reducing the cost of generating things, but not the cost of verifying them: . This explains why the first major AI products appeared in chat, image generation, and code assistance. Not because these were the hardest human problems, but because their outputs were relatively easy to inspect. A user can judge the tone of a message, look at an image, or run a test on a piece of code. […] The old automation boundary was routine versus non-routine work. The new boundary is increasingly measurable versus non-measurable work. The issue i

    judgment · Martin Fowler

  40. practice · Lenny's Newsletter ·

    🎙️ How I AI: GPT-6 Astra is a banger + Stripe’s AI playbook + Grok Bot vs. OpenClaw: why I replaced my entire agent stack

    Technical founder migrated six production agents from OpenClaw to Grok Bot, detailing setup, permission model, and personality preservation.

    Grok Bot vs. OpenClaw: why I replaced my entire agent stack Listen now on YouTube • Spotify • Apple Podcasts Brought to you by: WorkOS —Make your app enterprise-ready, with SSO, SCIM, RBAC, and more Hyperagent —Deploy fleets of agents that handle real work In this solo episode, Claire explains why she replaced her entire OpenClaw setup with Grok Bot. She walks through the bots she now uses to manage six inboxes, coordinate her family, review pull requests, monitor SOC 2 compliance, support customers, audit subscriptions, and shop for clothes. She also shares how she safely gives agents permiss

    ways of working · Claire Vo

  41. practice · Lenny's Newsletter ·

    Build your own company brain: the enterprise AI playbook from Stripe’s engineering team | Sharadh Krishnamurthy

    Stripe built internal AI agent Kai used by 10,000 weekly employees; governance through projects, safe data querying, skills platform lets any employee package workflows.

    Sharadh Krishnamurthy is an engineering manager at Stripe, where he helped build Kai, the company’s internal AI agent used by more than 10,000 employees every week. He’s worked across several of Stripe’s core infrastructure teams, including data and developer experience, which gives him a grounded, systems-level perspective on what it actually takes to make AI work at enterprise scale. He’s currently focused on the governance, skills, and infrastructure layers that let every Stripe employee use AI safely and effectively, regardless of their technical background. Listen or watch on YouTube , Sp

    adoption · Claire Vo

  42. practice · Nate B Jones ·

    Executive Briefing: How to Check Work You Didn't Watch Get Made

    As AI agents work unsupervised for days, executives must shift from directing work to judging finished results, creating new liability and trust challenges.

    The machine used to wait for you. You decided what it would do and when, it did that, and it stopped. That was true for as long as there have been computers. Agents started to change that in 2025, when they could take a goal and work toward it, and for the past year I’ve been setting goals for Sol, Opus 5, GLM 5.3, and others and judging what came back. Setting the goal and judging the result has been my job for a while. What changed this week, when OpenAI shipped GPT-6 Astra, is everything in between. Computing is now self-directed: you can state a goal briefly and imprecisely, the machine ru

    judgment

  43. research · Administrative Science Quarterly ·

    Cold Starts and Burning Desires: Reputation System Misalignment and Late-Career Transitions to Platform Work

    Late-career workers transitioning to platform freelancing struggle with reputation system misalignment between formal credentials and platform ratings.

    Workers face a cold start when entering new work contexts in which their reputations do not precede them. Cold starts are a particularly challenging problem for late-career workers who leave decades-long careers in formal organizations to pursue freelancing through online labor platforms. Unlike early- and mid-career freelancers, who are typically motivated by long-term career advancement, late-career workers are motivated to make work less central to their lives, although many still want and need to engage in paid labor in retirement. However, the question of how late-career workers navigate

    adoption · Paul Leonardi

  44. practice · Nate B Jones ·

    Seven sheets and 13 slides from the cheapest setting: the two-step guide and the exact prompts I use for Excel, PowerPoint, and Word.

    Run AI on cheap settings first to spot problems, then upgrade to expensive models only for judgment-requiring fixes.

    Fable 5.1 has changed how I think people should use frontier models for knowledge work. I asked it to research GoPro’s pending merger with Starman Optical, build a post-close discounted cash flow model in Excel, and turn the analysis into a PowerPoint. On Low, from one short prompt, it produced a seven-sheet workbook with working formulas, scenarios that actually moved the valuation, and a 13-slide deck I could use. Start it on Low and let it build the workbook, the deck, or the draft. Then open what it made, and move up to High or Extra only for the part that needs judgment. The reason this w

    ways of working

  45. research · Information Systems Research ·

    Enhancing AI Use: How Complementary System Information Drives Delegation Frequency and Effectiveness

    Providing both AI confidence scores and outcome feedback together increases how often workers delegate to AI and improves delegation quality, boosting combined performance.

    For a collaboration between humans and artificial intelligence (AI) to be fruitful, tasks should be allocated based on their complementary capabilities. Prior research shows that when humans are responsible for allocating tasks between themselves and an AI through delegation, they often delegate too infrequently or delegate the wrong tasks, preventing complementary performance gains.We study how different types of AI system information affect both delegation frequency and delegation effectiveness, which capture the extent to which humans can leverage existing complementarities with AI. Specifi

    judgment

  46. practice · Hacker News ·

    Porting my 1993 Amiga game to Godot, with an LLM reading the 68000 assembly

    Developer used an LLM to port 33-year-old 68000 assembly code to Godot, achieving a working game in one evening.

    These are my notes from porting my Amiga game, which I originally built in Baghdad in 1993 in MC68000 assembly, to Godot, using Claude Fable 5 during last July holiday. It took an evening! Getting the feel right and shipping it took a few more weekends and evenings. I spent the last few weeks analyzing what Claude did, feeding it my 33 years of memory of how I built the game, my notes and the git repos. It wrote the first draft of the article, and I edited line by line over a week. The screenshots of my 1993 map editor is the first I have run it since then. The one thing I never verified mysel

    ways of working

  47. research · HBS AI Institute ·

    Think You’re a Good Leader? Try Leading an AI Team

    An AI-based assessment tool tracked leaders' true impact on human teams by controlling for team member ability.

    Researchers found that an AI-based assessment closely tracked leaders’ impact on human teams. Listen to this article: Judging a leader with a successful team is like judging the conductor of a great orchestra: the outcome reflects both the person directing the group and the abilities of everyone around them. Measuring leadership therefore requires a way […] The post Think You’re a Good Leader? Try Leading an AI Team appeared first on Harvard Business School AI Institute .

    management org

  48. research · ACM Transactions on Software Engineering and Methodology ·

    Collaborative Knowledge Distillation and Reinforcement Learning for Automated Ticket Triage in Large-Scale Production Systems

    Automated ticket triage system deployed at ByteDance reduced average triage time to 15 seconds using knowledge distillation and reinforcement learning.

    In large-scale enterprise environments, growing system complexity makes failures inevitable, threatening business continuity and customer satisfaction. To maintain system stability, efficient ticket triage is crucial for timely incident resolution. However, it remains a knowledge-intensive task requiring substantial domain expertise and detailed analysis, revealing the limitations of traditional rule-based and text-based methods. Recent advances in large language models (LLMs) offer new possibilities for automating triage through their remarkable reasoning and language capabilities. Yet, in in

    productivity

  49. research · npj Digital Medicine ·

    Differences in tone of AI and care team responses to patient messages by patient demographics

    AI-drafted patient messages showed demographic biases in tone, with lower politeness and positive affect for Hispanic, non-English, female, and lower-income patients.

    The use of artificial intelligence (AI) to draft responses to patient portal messages has been proposed to reduce provider in-basket burden. However, little is known about its effects on patient–provider communication. In this retrospective observational study, we evaluated demographic differences in the tone of AI-generated draft replies (AI-GDRs) and care team responses to patient messages. Our study included 12,202 message triads comprising patient messages, AI-GDRs, and care team responses from three internal and family medicine practices in New York City. We found differences in tone acro

    judgment

  50. practice · martinfowler.com ·

    An Accidental Blackboard

    A team using agentic engineering discovered agents self-organized into a blackboard coordination pattern, emergently stored in git.

    Giles Edwards-Alexander reports that during an experiment to see how productive a team could be using fully agentic engineering practices, the team accidentally prompted the agents into creating a blackboard coordination system inside the git repository. more…

    ways of working · Martin Fowler

  51. practice · martinfowler.com ·

    Maybe We Shouldn't Be Reviewing All This Code

    Code review may not be the right tool for AI-era workflows; consider what problems it actually solves before automating it away.

    TL;DR Or, perhaps the problem isn't that AI has broken code review, maybe it’s that we've been using code review to solve the wrong problems I was on a panel recently with Brian Houck from DX at Code Remix, hosted by Moderne. It was one of the more interesting panels I’ve done, largely because we disagreed. As my colleague Martin Fowler says, panels are much more interesting when people disagree and both sides have a good argument. Brian and I definitely did. Brian has since written a thoughtful piece called What are code reviews even for? He is clearly passionate about his position, and I am

    ways of working · Martin Fowler

  52. practice · Nate B Jones ·

    Your AI provider can change the deal on you. Here's the 5-prompt audit I run to stay ready.

    A five-prompt audit to test whether your AI workflows are portable across providers or locked in.

    November 12 is the day OpenAI’s models are set to stop working directly inside Cursor. Nobody using Cursor did anything to cause that. SpaceX bought the company, OpenAI decided it couldn’t be confident in the new owner, and the model leaves. Monday, over on YouTube , I asked which AI future you were betting on: the one you rent from a frontier lab, or the one you own. These aren’t separate stories. Three companies just answered that question with their money. And, if you watched today’s video, you already know the map: OpenAI is trying to own more of the AI system, NVIDIA is trying to sell the

    ways of working

  53. practice · How I AI ·

    Grok Bot vs. OpenClaw: How I replaced my entire agent stack

    A practitioner running 30 agents across work and life describes specific bots, migration steps, and unexpected outcomes like helpdesk agents earning five-star reviews.

    I’m running about 30 active agents at any given moment, and in this episode I break down my full Grok Bot setup: what it is, how it compares to OpenClaw, and the nine bots I’ve built for work and my personal life. We go deep on Chief (my chief-of-staff bot sweeping six inboxes and multiple Slack workspaces), TradBot (the family agent that prints a kitchen-table newspaper for my kids), two engineering bots handling my PR queue and SOC 2 compliance monitoring, Holly Helpdesk, and a handful of personal bots I didn’t expect to actually love. I also walk through how I migrated everything from OpenC

    ways of working · Claire Vo

  54. research · arXiv ·

    Meeting the Coming Wave: The Emerging Politics of AI and Work across 33 Parliaments

    Analysis of 1.5 million parliamentary speeches shows AI-work politics focuses on enablement (55%) and regulation (22%), not worker compensation (2.3%), with left-right divides over governing trajectory.

    A new politics of artificial intelligence and work is taking shape across party systems, but comparative politics has yet to map it. Using 1,514,950 parliamentary speeches from 33 parliaments (2023-2026), we show this politics follows a different logic than political economy expects. Research anticipates that technological disruption generates demands for compensation; instead, compensation accounts for just 2.3% of response-frame mentions, while enablement and investment dominate (55.2%), regulation and restriction follow (21.8%), and training (20.6%) appears at similar rates across families.

    policy

  55. research · arXiv ·

    Knowing Is Not Enough: Information Retrievability as a Precondition to Effective LLM Oversight

    Self-generated explanations and retrieval cues help customer-facing employees detect LLM errors better and sustain oversight as use becomes routine.

    Large language models (LLMs) are increasingly embedded in organizational work, yet their errors often pass human review. Prior research locates such failures in users' capability to review LLM output or their engagement in doing so. We develop an alternative, retrieval-based account of human oversight and posit that error detection is more effective when oversight-relevant information is accessible to users at the moment of review. Across two randomized lab-in-the-field experiments with 640 customer-facing employees, we show that self-generated explanations improve error detection and strength

    judgment

  56. research · NBER ·

    Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting

    A pre-registered RCT with 133 patent lawyers over three months tests whether AI drafting assistance builds or erodes professional expertise.

    Whether AI assistance builds or erodes professional expertise is unsettled. In a pre-registered three-month randomized controlled trial, we gave 133 practicing patent lawyers at eleven U.S. intellectual property law firms access to a custom AI drafting assistant and measured both their performance (David Autor , Tanya Rodchenko , Josh Martin , Zanna Iscenko , Scott Strand , David Pearl , Melissa Ferere)

    jobs skills · David Autor

  57. research · NBER ·

    The Macroeconomic Effect of AI: Sizing the Software Engineering Channel

    Measures AI's macroeconomic effect via software engineering productivity using real-time stock market data.

    We measure how artificial intelligence (AI) affects the economy through its impact on software engineering productivity. We use information from financial markets to develop a forward-looking measure that is available in real time. We estimate the sensitivity of each firms stock return to an AI (Alex Blumenfeld , Jonathon Hazell , Chen Lian , Andreas Schaab)

    productivity

  58. research · NBER ·

    The Early Impacts of AI on Employment among Recent College Graduates

    First estimates of AI's effect on unemployment rates of recent college graduates using CPS microdata.

    The impact of AI on the employment prospects of recent college graduates is hotly debated with no consensus on the magnitude of impacts nor even the timing of those potential impacts. Using CPS microdata, we provide the first estimates of the effects of AI on the unemployment of recent college (Robert W. Fairlie , Jane Wu)

    jobs skills

  59. research · Research Policy ·

    Minds and machines: Rethinking absorptive capacity in the age of artificial intelligence

    AI changes how firms absorb and learn from knowledge; traditional absorptive capacity theory needs revision to account for AI as a learning agent.

    The absorptive capacity construct derives its functions, dimensions, and internal relationships from the implicit assumption that humans are the only learning agents underlying knowledge absorption. I posit that the emergence of artificial intelligence (AI) challenges this assumption and requires a re-examination of the construct. Through a theory-building abductive study combining conceptual mapping with qualitative interviews, this work investigates the complex relationship between AI and knowledge absorption. It theorizes new dimensions of absorptive capacity pertaining to AI-driven process

    adoption

  60. practice · Addy Osmani ·

    Agentic Skill Decay

    Deliberate practice habits needed to maintain expertise when AI agents handle routine coding work.

    Mastery still comes from doing the reps. Before agents, I got my reps as part of writing code: try different approaches out, debug what went wrong, review other’s code, read a lot. Agents can skip much of that work, so building your reps has to be deliberate . If I was new to the industry, I’d try to form a hypothesis before prompting. Ask “why” a lot, read the diffs, try to predict what might fail. Occasionally try to work through the problem myself manually. Your AI agent has database access. Can you tell what it did? Give an agent a Postgres login and it looks like any other role - a shared

    jobs skills · Addy Osmani

  61. research · MIT Sloan Management Review ·

    Three Things to Know About Customer Resistance to AI

    Summarizes three studies on conditions under which customers avoid or accept AI chatbots replacing human service reps.

    Microsoft Copilot/Unsplash Companies are betting that AI chatbots will deliver faster and cheaper customer service. But if you’ve ever tried to circumvent a chatbot and get to a human, you’re not alone. Here’s what three recent studies discovered about when customers will and won’t let AI do a human’s job. 1. Customers avoid chatbots for […]

    adoption

  62. practice · How I AI ·

    How I turned Claude into a self-improving PM assistant | Daniel Blum (PM, Melio)

    PM built self-improving Claude agent that automates weekly prep, learns company jargon, and scales personalized setup across team in 15 minutes.

    Daniel Blum is a product manager at Melio, a B2B payments company, and one of the most systematic thinkers I’ve had on the show when it comes to personal AI infrastructure. He’s spent the past year building a Claude- and Cowork-based productivity system that manages his Notion board, processes his Slack and email, and runs self-improvement loops every week without needing to be prompted. Beyond his own workflow, Daniel built and scaled a “Workstation” onboarding plugin that gets any Melio employee up and running with a personalized Claude setup in about 15 minutes. What you’ll learn: Why Danie

    ways of working · Claire Vo

  63. practice · One Useful Thing ·

    Agency and Agents

    AI systems now act autonomously rather than waiting for instructions, raising questions about human versus machine agency in work.

    Agency is the initiative to act. Increasingly, it is going to determine what happens next with AI, and whether that is good or bad for us. But whose agency? Human agency, the willingness to push, experiment and act without waiting for instructions, seems increasingly important to getting value out of AI, and I have a longer post on that coming soon. But this post is about the agency of AI, and how the choices we make about how to use it (or constrain it) will shape all of our futures. For much of the last few years, the AI would sit in a chat window until you asked it for something. Even when

    management org · Ethan Mollick

  64. practice · Sangeet Paul Choudary ·

    How agents transform the workflow, the organization, and the competitive landscape

    Agents reshape B2B procurement workflows, internal decision architecture, and competitive dynamics between buyers, sellers, and software providers.

    Much of today’s discussion about agentic commerce focuses on whether shopping will move from e-commerce websites into interfaces such as ChatGPT. But in consumer commerce, much of the buying journey has already been solved reasonably well by the browse-and-click paradigm. Comparing options was the last remaining bottleneck that ChatGPT now, often, solves better than the traditional listing search engine. B2B procurement is different. Much of the buying journey has never been effectively digitized because the work cannot be reduced to browsing a catalogue and clicking through a predetermined ro

    management org · Sangeet Paul Choudary

  65. practice · John Warner ·

    Maintaining Your Ignorance

    Writer describes how assuming AI-generated messages and maintaining defensive skepticism shapes decisions about engagement and trust.

    At this point, as far as you know, this is your only chance to subscribe on this page. Six or so months ago I received a message through my website from a purported reader of my novel, The Funny Man , who said he’d loved the book, thought it was “sad and hilarious,” and wanted to know if I had any more full-length fiction on the horizon. My immediate and firm assumption was that this was an AI-generated come-on that, if responded to, would result in some kind of ask for money to promote the book to a reading group or to “increase its online visibility.” I’d received literally dozens of these e

    judgment · John Warner

  66. practice · Nate B Jones ·

    Codex, Grok and Claude all agree, and you still don't know if they're right. The guide I use to decide.

    Deliberately add friction to AI workflows to preserve decision-making skills and avoid productivity-masking deskilling.

    I use AI all day, every day, and I’d like to believe it has made me smarter. If simple exposure caused brain rot, I should be a cautionary tale. By smarter, I mean my own thinking feels sharper and my taste is easier to articulate. I find the edge of what a new system can do before its failure patterns cost me real time. A lot of people worry that brain rot is a byproduct of laziness. The kind I fear looks like extraordinary productivity. More pages, more plans, more code, more finished-looking work, all arriving faster than ever. The model forms the first opinion, writes the plan, resolves th

    worker experience

  67. research · arXiv ·

    Sophistication in GenAI Use: Field Evidence from a Large Firm

    Senior employees use AI more sophisticatedly; sophistication varies by function but does not improve with training or time.

    We study how sophistication in generative AI (genAI) use varies among the back-office workforce of a large firm. Using proprietary data, we observe 713,564 employee prompts and their corresponding large language model responses from nearly 4,000 back-office employees across 15 functional areas over eight months in 2025. We document three main findings. First, senior employees exhibit more sophisticated genAI use, consistent with domain expertise complementing genAI capabilities. Second, sophistication varies considerably across functions and is highest in Strategy, Digital Innovation, and Proj

    adoption

  68. practice · Addy Osmani ·

    Audit your Agent files

    Agent instruction files degrade over time; audit them regularly by running /doctor, reviewing memory, and removing instructions that no longer add value.

    TL;DR: Your coding agent’s configuration has a half-life. Models improve, harnesses add capabilities, codebases change, and the instructions we wrote for an older version stay behind. Recent research finds inconsistent value from personalized skills. I now run Claude’s /doctor every few weeks, review memory separately, and ask each instruction to earn its place again. I feel like there’s been a lot of confusion about skill files and what to do with your CLAUDE.md and AGENTS.md files, especially as I’ve been reading developer discourse on Twitter over the last few months. People have been sayin

    ways of working · Addy Osmani

  69. practice · martinfowler.com ·

    Making Your Data Ready for Agentic AI

    How to prepare organizational data so AI agents can access accurate, trustworthy information for autonomous work.

    Lots of organizations are excited about what AI can do to streamline their processes, save money, and juice margins. But AI's capabilities are founded on the data that AI accesses, and for many organizations that foundation is little more than sand. Pramod Sadalage and Prem Chandrasekaran write about how to build a reliable foundation of data that can be accurate and trusted. more…

    adoption · Martin Fowler

  70. research · arXiv ·

    Monocultural Biases: Correlated biases in large language models lead to unequal systemic exclusion rates in hiring

    Post-trained LLMs show correlated hiring biases across models, increasing systemic exclusion rates from 5.6% to 17.3%, with age discrimination worsening after training.

    Employers are increasingly using large language models (LLMs) to automate their hiring process. This paper investigates the risk of monocultural biases, in which the widespread deployment of large language models homogenizes biases across the labor market, leading to greater systemic exclusion for certain demographic groups. For ten LLMs, we measure hiring biases across their base and post-trained versions to identify which stage, pre-training or post-training, lead to monocultural biases. We find that, compared to their base models, post-trained models are 3.6% less likely to callback older a

    jobs skills · Maria del Rio-Chanona

  71. practice · Nate B Jones ·

    You bought the agent to get time back. Here is why your calendar filled up instead (+ the five prompts that fix it.)

    Agent work creates invisible overhead: allocation, specification, evaluation, intervention, coordination, recovery that fills calendars despite time-save promises.

    I can end a day carrying work that the product, the team, and the budget have never named. The strange part of running agents is that I start more work than I can inspect. I spend the day moving between outputs that each need a decision. None of it shows up anywhere. The dashboards report tokens, run counts, and time saved during execution, and not one of them measures the thing that actually filled the day. That work has a shape. Allocation, specification, evaluation, intervention, coordination, recovery. It takes judgment and it carries accountability. Some days I find it exhausting. Agent f

    worker experience

  72. research · arXiv ·

    Are Concept Bottleneck Models Effective as Decision-Support Systems?

    Large user studies show concept-based AI explanations improve human-AI team accuracy, but only when tasks are difficult, concepts are clear, and users actively interact.

    Concept Bottleneck Models (CBMs) are interpretable-by-design neural networks that detect human-understandable concepts from the input and use them to generate predictions. By allowing users to inspect the concepts underlying a prediction and explore how predictions change under alternative concept configurations, CBMs have emerged as one of the most prominent approaches to supporting human-AI collaboration. However, user studies investigating their actual effectiveness as decision-support systems remain limited. We present two large-scale user studies (N participants = 705, N observations = 6,

    judgment

  73. practice · GitHub Blog ·

    How to evaluate LLMs before production

    GitHub reduced false positives in secret scanning by moving from benchmark evaluation to production-realistic testing with safety constraints.

    A language model can perform well on a clean benchmark and still struggle with the cases that matter in production. Benchmarks and curated datasets are useful when prototyping an LLM-based system. They help teams compare models, test an initial prompt, and determine whether an idea is technically plausible. But as a system moves closer to production, the evaluation problem changes. Real inputs are often ambiguous. Labels may be inconsistent. Important context may be missing or truncated. The evaluation set may not reflect the production distribution. Edge cases that rarely appear in benchmarks

    judgment

  74. practice · The Pragmatic Engineer ·

    Why Ramp built its own in-house coding agent, Inspect

    Ramp built Inspect, an internal coding agent running on remote sandboxes with access to internal data, to run multiple agents in parallel and integrate with their development workflow.

    At a select few tech companies, they write most of their code with their own, custom-built, internal AI coding agents. This is different from most of the industry which uses AI coding agents and harnesses like Codex, Claude Code, Cursor, OpenCode, GitHub Copilot, etc. At Ramp, their own version is called Inspect , while at Block it’s Goose (open source), at Stripe it’s Minions , and River at Shopify. But why not just use what frontier labs and coding harness AI startups already offer; why take the time and effort? We reached out to Ramp, a fintech company big on building its internal AI infras

    ways of working · Gergely Orosz

  75. research · EU Joint Research Centre ·

    A rising tide: Revisiting the occupational impact of AI in the generative era

    EU researchers update occupational exposure estimates for generative AI, revising earlier automation risk rankings across jobs.

    A rising tide: Revisiting the occupational impact of AI in the generative era

    jobs skills

  76. research · Strategic Management Journal ·

    The impact of generative artificial intelligence on innovation: Evidence from software products

    ChatGPT release increased new software products by 34% but reduced improvements to existing ones by 20%, shifting developer effort toward quantity over novelty.

    Abstract Research Summary We study the impact of generative artificial intelligence (GAI) tools on product‐level innovation outcomes in the context of software products. Specifically, we illustrate how GAI can alter the direction of innovation by shifting the activities of developers away from generational innovation and toward original innovation, which may be new to the market but not necessarily more novel than previous innovations. We argue that this shift is driven by GAI's ability to facilitate tasks in both the ideation and implementation of software products, which enables some develop

    adoption

  77. practice · How I AI ·

    I spent $20,000 on Devin in a month. Here’s what I learned | Ryan Carson (solo founder)

    Solo founder manages 15 concurrent AI agents with paper-based system; replaces QA team using LAN PR skill and Watchdog playbook for customer success.

    Ryan Carson is a five-time founder and the current solo founder of Untangle, a B2B SaaS platform for family law firms. Before Untangle, he co-founded Treehouse, an online coding education platform, and has spent the better part of two decades building and leading tech companies. He’s active on X, where he shares his solo founder journey in real time, including what he actually spends on AI tools each month. What you’ll learn: Why Ryan manages 15 concurrent Devin agents with a folder system and a piece of paper, not a dashboard The Watchdog playbook: what he built to replace a customer success

    ways of working · Claire Vo

  78. practice · Claude Blog ·

    How an Anthropic field marketer uses Claude Code to send weekly personalized updates to every sales rep

    Field marketer uses Claude Code to generate personalized weekly updates for each sales rep, automating segmentation and customization at scale.

    How an Anthropic field marketer uses Claude Code to send weekly personalized updates to every sales rep

    ways of working

  79. research · Information Systems Research ·

    The Impact of Generative AI on Collaborative Open-Source Software Development: Evidence from GitHub Copilot

    GitHub Copilot increases code contributions but raises coordination costs; benefits concentrate among core developers while peripheral contributors face integration delays.

    Generative AI holds great promise for reshaping software development. We examine the impact of GitHub Copilot, a generative AI pair programmer, on collaborative open-source software (OSS) development, where developers voluntarily contribute and collaborate on projects. Using GitHub’s proprietary Copilot usage data combined with public OSS data from GitHub, we find that Copilot increases project-level code contributions by expanding developer participation and increasing code contributed by individual developers. However, these gains lead to increase in time to coordinate, review and integrate

    teams

  80. practice · Sangeet Paul Choudary ·

    Strategy in the age of AI - Eight points beyond the obvious

    AI shifts competitive advantage from modularity to owning feedback loops connecting activities across an organization.

    What happens when AI changes not only how companies compete, but the very game they are competing in? In this conversation, Rita McGrath and I - guided by the always-insightful Aidan McCullen - bring together complementary ideas from the books we have written and those we are currently working on. It promises to be a treat: a wide-ranging exploration of what strategy becomes when the assumptions it rests on begin to shift. Rethinking strategy Our central argument is that viewing AI through the traditional mechanisation/automation lens traps us in looking for operational benefits (much of the d

    management org · Sangeet Paul Choudary

  81. research · Exponential View ·

    🔮 Why one AI is better than four #598

    Multi-agent AI systems fail at hidden-profile problems, choosing wrong answers 64-83% of the time versus single agents choosing correctly.

    Good morning! We are looking for an outstanding economist to join us as an AI Economy Research Fellow . If you know someone we should speak to, send them our way. Great minds think (a little too much) alike A few months ago, we (alongside ) looked at whether AI is immune to groupthink . The answer was no. Blending several models’ answers kept about a quarter of the good ideas that had come from a single model. This is called the hidden-profile problem: when groups discuss what everyone already knows and don’t get to the knowledge that only one member holds. Anthropic has now run that classic e

    judgment · Azeem Azhar

  82. research · ACM Transactions on Software Engineering and Methodology ·

    What Needs Attention? Prioritizing Drivers of Developers’ Trust and Adoption of Generative AI

    Survey of 238 developers at GitHub and Microsoft identifies system quality, output quality, functional value, and goal maintenance as drivers of trust and adoption of genAI tools.

    Generative AI (genAI) tools promise productivity gains, yet developers still struggle to determine when to trust and effectively integrate these tools into their everyday work. Moreover, genAI can be exclusionary, failing to adequately support developers across individual differences. One such difference is cognitive style , which can shape how developers engage with genAI (e.g., risk-averse developers may gate outputs behind tests, whereas risk-tolerant ones may prototype directly and address issues post hoc). When tools fail to accommodate these differences, they can create additional usabil

    adoption

  83. practice · Addy Osmani ·

    Human judgment doesn't leave the software factory. It relocates.

    Human judgment relocates from writing code to deciding intent, design, quality bars, and shipping gates in AI-assisted software factories.

    A software factory is a repeatable loop around software work. If you’re building a software factory, code good enough to ship still needs human taste and ownership. We’ll discuss this including whether you need a factory just yet. If so: You’ll likely need humans in the loop upfront for deciding on product intent, system design (if you care) and your quality bar. Do review code (lights-on factory) but be intentional with where it’s needed the most. I’ve found you want to watch out for where automated back-pressure breaks. Or where maintainability trade-offs need to be made. Aim for quality che

    judgment · Addy Osmani

  84. practice · The Pragmatic Engineer ·

    The Pulse: We need to talk about migrations with AI

    AI completed a two-week test framework migration at Asana that would have been indefinitely postponed without it.

    The Pulse is a series covering events, insights, and trends within Big Tech and startups. Today, we cover: More on the “great engineering leader career break.” The industry is changing fast, and the VPE and CTO roles also need to adapt. And don’t forget that these are the roles from which you can drive change that reorganizes engineering in ways that work better. We need to talk about migrations with AI. Asana needed to migrate off testing framework Enzyme, but it meant doing a massive rewrite of test cases. With AI, the project was completed in two weeks: without AI, this work would surely ha

    productivity · Gergely Orosz

  85. research · MIT Sloan Management Review ·

    Algorithms Trap Us in the Familiar. Can They Also Spark Breakthroughs?

    Standard recommendation algorithms push team members toward familiar popular information, causing convergent ideas, unless redesigned to surface diverse content.

    Most digital tools are designed for efficiency, not exploration — and that can stymie innovation. New research shows that standard algorithms channel team members toward popular, familiar information, leading them to independently converge on the same ideas. But the researchers uncovered a surprisingly simple fix: By using algorithms designed to surface diverse, uncommon information, domain experts can avoid “ideation bubbles” and generate ideas across entirely new solution spaces.

    teams

  86. research · JAMA Network Open ·

    An Electronic Health Record–Integrated, Large Language Model–Powered Tool to Triage Surgical Patients

    LLM-based triage tool achieved 86% sensitivity and 78% specificity for identifying surgical patients needing comanagement, with physician review.

    Importance: Surgical comanagement (SCM) is an evidence-based care model in which hospitalists jointly manage medically complex perioperative patients alongside surgical teams. Despite its clinical and financial value, effective use of SCM is limited by the need to manually identify eligible patients; large language models (LLMs) are increasingly prevalent in clinical workflows and could be useful in selecting patients for SCM. Objective: To assess whether SCM eligibility triage can be automated. Design, Setting, and Participants: This prospective, unblinded quality improvement study was conduc

    adoption

  87. research · arXiv ·

    Who Delegates to AI? Evidence from Agent Configurations in Github

    Occupations where workers delegate tasks to AI agents differ markedly from those vulnerable to pre-AI automation; adoption patterns vary by education and wage level.

    A growing body of literature measures the extent to which occupations are exposed to AI, yet existing measures capture where AI could perform tasks rather than whether workers have actually adopted it. We introduce a distinct tier of exposure, delegated exposure, which records whether a worker has committed a task to AI by embedding it into a structured workflow. We operationalize this concept through the Agentic Adoption Index (AAI), measuring how closely an occupation's tasks align with the agentic routines that practitioners have built and shared. Using semantic embeddings of roughly 888,00

    adoption

  88. research · arXiv ·

    CentaurBench: Benchmarking LLM Capabilities on Augmenting vs. Automating Real-World Work Tasks

    LLM models that best automate tasks often fail to best augment weaker agents; automation and assistance capability rankings are only modestly correlated.

    Most LLM benchmarks rank models on their ability to automate work tasks. In practice, however, models are often used to assist other (human or LLM) agents. The question that drives model selection is therefore not only which model produces the best output, but which model most improves the work of another (weaker) agent. We introduce a unified framework that evaluates the capability of models to automate and augment another agent's performance. Across seven economically grounded real-world tasks, an assistant model writes assistance text for a standardized lower-capacity worker model, which pr

    adoption

  89. research · Management Science ·

    Beyond the Black Box: Unraveling the Role of Explainability in Human-Artificial Intelligence Collaboration

    Explainable AI reduces cognitive burden at low levels but increases fatigue at high levels, with effects varying by decision complexity and time pressure.

    Explainable artificial intelligence (AI) models have been proposed to mitigate overreliance and underreliance on AI, which reduce the effectiveness of human-AI collaborative tools. Yet, empirical evidence is mixed, and the impact of explainable AI on the cognitive effort and fatigue of a decision maker (DM) is often overlooked. This paper offers a theoretical perspective on these issues. We develop an analytical model that incorporates the defining features of human and machine intelligence, capturing the limited but flexible nature of human cognition with imperfect machine recommendations. Cr

    judgment

  90. research · Organization Science ·

    Filtering the labor pool: How digitalization redistributes hiring costs

    Randomized experiment shows hiring filters reduce interviews by 3.4% overall but screen out 28.9% fewer low-quality and 36.9% fewer high-quality applicants, raising wages 2.5%.

    I theorize that candidate filters, a core technology in digital labor markets considered central to cost-effective matching, redistributes costs from the search-and-screening stage to the job-offer stage of hiring. Filters cause employers to focus on applicants who display common preferred characteristics. At the search-and-screening stage, filters reduce the number of applicants interviewed and alter the composition of the interviewed pool by removing both lower- and higher-quality candidates. At the job-offer stage, filters lead employers to make job offers to candidates with easily observab

    jobs skills

  91. practice · The Pragmatic Engineer ·

    Headed for the Exit: the Great Engineering Leader Career Break

    Engineering leaders are quitting high-status roles at unprecedented rates, citing AI disruption, burnout and role devaluation as key factors.

    In my ~20 years in this industry, I’ve not seen as many capable engineering leaders opting out or taking prolonged breaks as now, with some high-ranking engineering leaders – CTOs, VPs of Engineering, heads of engineering, etc. – quitting their high-status roles and departing, if not into the sunset, then at least with nothing lined up. To find out what might be behind this spate of sign-outs, I talked with almost 20 engineering leaders currently on a career break – or seriously considering one – and they let me into their personal reasons for deciding to jam the brakes on their careers. Thank

    jobs skills · Gergely Orosz

  92. practice · Claude Blog ·

    Claude on call: How Claude Tag serves as Anthropic’s first responder for CI/CD failures

    AI agent handles initial triage and response for CI/CD pipeline failures, reducing manual oncall burden.

    Claude on call: How Claude Tag serves as Anthropic’s first responder for CI/CD failures

    productivity

  93. practice · How I AI ·

    How a solo founder used Codex and ChatGPT to launch a fashion brand without engineers | Yana Welinder

    Solo founder built fashion brand from sketch to live e-commerce using AI as technical co-founder, with detailed fashion prompt as technical spec.

    Yana Welinder is the solo founder of Yana Bana, an AI-native fashion brand built with AI as her technical co-founder, starting from hand-drawn sketches and ending with runway photos, CAD files for 3D printing, and a live Stripe-connected pre-order site—no engineers required. A former product leader, she brings an operator’s rigor to her creative process: her “fashion prompt” is a detailed spec covering silhouette, volume, fabric behavior, movement, and sound, and watching her use Codex plus computer use to navigate 3D design software that’s entirely new to her is a clarifying demo of what toda

    ways of working · Claire Vo

  94. research · arXiv ·

    Adoption of Generative AI in the Workplace: Increasing and Shifting the Balance of Productivity and Communication Activity

    AI adoption correlates with 21.2% more productivity actions and 7.1% more communication actions among heavy users, with a shift toward documentation work.

    Generative AI is transforming the workplace by augmenting and automating cognitive tasks, reshaping how organizations work and innovate while raising questions about workplace inequality and the future of work. Despite rapid adoption, empirical evidence on how these tools alter work practices and generate productivity gains remains limited. We examine how AI use affects the quantity and nature of information work using digital trace data from the Microsoft M365 application suite across multiple large international companies. Specifically, we study how generative AI adoption shifts the balance

    adoption · Siddharth Suri · Scott Counts

  95. research · Government Information Quarterly ·

    Working with semantic machines: The role of semantic reconciliation in public sector data governance

    AI recruiting systems require continuous organizational work to reconcile standardized vendor outputs with local context and occupational knowledge across distributed government teams.

    Local governments are beginning to use artificial intelligence (AI) in internal administrative work, including recruitment. In a decentralized municipality, however, a standardized AI assessment must travel between a technology vendor, central HR specialists, recruiters, and hiring managers responsible for different services. This creates a tension between representational portability and contextual adequacy. Based on a qualitative case study of a large Swedish municipality, we examine an AI interviewing system that translated candidates' open-ended responses into competency indicators and nar

    adoption

  96. practice · Addy Osmani ·

    Practical Loop Engineering

    Running five to ten agents in parallel with different oversight levels, using loop engineering for self-correcting cycles.

    The way that I typically work is I have anywhere between five and ten agents working at the same time in parallel. There are going to be some tasks that I’m very happy to delegate fully to agents, as long as I have a very clear idea of the stopping conditions and the constraints around them. And then there are going to be some tasks where I am going to want to keep a closer eye and code-review what the agent is doing. Now within that, you’ve probably heard about loop engineering . I talked about it a couple of months ago when I wrote a big blog post about it. A loop is an autonomous, self-corr

    ways of working · Addy Osmani

  97. practice · Kent Beck ·

    Baking a Model

    Kent Beck uses baking as an analogy to explain model construction mechanics and sensitivity to initial conditions.

    I remember walking to the bus from high school, staring at a Motorola 6800 instruction set manual. I didn’t really understand what I was looking at—boolean expressions, instruction encodings, timing tables—but I was obsessively fascinated by the mechanism of it all. Here was this complicated machine where if I understood it I would have power & control. I feel the same way about AI models right now. I don’t claim to understand the details, not yet, but I’m fascinated by the mechanism of it all. I’m interested in both: How models work but also, The machinery that makes a model. It’s this latter

    ways of working · Kent Beck

  98. research · HBS AI Institute ·

    How AI Agents Are Changing the Way We Work

    A working paper finds agentic AI expands both the efficiency and the scope of tasks knowledge workers can complete.

    New research shows the next wave of AI expands both productivity and the scope of work. Listen to this article: For businesses, the most important question about agentic AI is whether it actually changes what work people can get done. The new working paper, “How AI Agents Reshape Knowledge Work: Autonomy, Efficiency, and Scope,” begins […] The post How AI Agents Are Changing the Way We Work appeared first on Harvard Business School AI Institute .

    productivity

  99. research · arXiv ·

    Applied and Filtered: An End-to-End Algorithmic Fairness Audit of A Public Employment Agency

    End-to-end audit of a public employment agency's AI hiring system reveals disparities by age, salary level, and gender identity masked by aggregate parity metrics.

    Algorithmic fairness evaluation commonly assesses AI systems as bounded technical components, abstracting away the organizational context in which they operate. We present, to our knowledge, the first independent end-to-end fairness audit of a semi-automated hiring system operated by Barcelona Activa, a public employment agency using the third-party TalentClue platform for candidate search and shortlisting. We analyze approximately 497,000 candidate-vacancy pipeline entries from September 2017 to September 2022, covering seven pipeline stages that span automated processing, human discretion, c

    adoption

  100. research · npj Digital Medicine ·

    Beyond expertise: exploring behavioral personas in AI-assisted rare renal cancer diagnosis

    Pathologists using AI for rare cancer diagnosis showed different behavioral patterns: some followed AI recommendations often (automation bias), others questioned them more, with gains highest among less experienced doctors.

    Abstract Fumarate Hydratase-deficient Renal Cell Carcinoma (FHdRCC) is a rare (< 0.4% of RCCs), aggressive subtype with significant morphological overlap, posing diagnostic challenges. While artificial intelligence (AI) shows promise in common cancer diagnostics, its impact on pathologist decision-making in rare diseases—particularly concerning automation bias—remains poorly understood. We developed a deep learning model to classify FHdRCC. We conducted a crossover reader study with 21 pathologists (7 genitourinary (GU) specialists, 7 non-GU specialists, 7 residents) diagnosing 30 challenging

    judgment