What do coding agents change about software work?

The very large productivity gains from agents require working with them directly; putting another human in between is too slow to preserve the gain.

1. David Heinemeier Hansson 26 Aug 2026 The revelation I've had working on Omarchy the last three months is that to get that magical 10X, 100X, in a few rare cases, 1000X productivity boost, you have to interact with the agents directly, and you cannot intermediate that bandwidth with another human because it's simply too slow.

   Creator of Ruby on Rails, CTO of 37signals

   Their wordsThe revelation I've had working on Omarchy the last three months is that to get that magical 10X, 100X, in a few rare cases, 1000X productivity boost, you have to interact with the agents directly, and you cannot intermediate that bandwidth with another human because it's simply too slow.

   spoken · official transcript hear it at 0:19:14 · all korrents from this recording

   ↗DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501youtube.com

Agent-written pull requests already outclass those of the median human contributor to an open source project.

1. David Heinemeier Hansson 26 Aug 2026 But it actually means that if you take the median programmer and their pull requests towards an average open source project, they're already getting outclassed by agents. I would rather get an agent-written pull request to one of my projects than I'd get one written by a human, and it's not just because the quality's better. It's also because I feel a lot less bad if I just reject it.

   Creator of Ruby on Rails, CTO of 37signals

   Their wordsBut it actually means that if you take the median programmer and their pull requests towards an average open source project, they're already getting outclassed by agents. I would rather get an agent-written pull request to one of my projects than I'd get one written by a human, and it's not just because the quality's better. It's also because I feel a lot less bad if I just reject it.

   spoken · official transcript hear it at 0:31:58 · all korrents from this recording

   ↗DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501youtube.com

There is no accumulated advantage in agentic coding: someone who ignored the whole of the past year could catch up to the frontier in two weeks.

Filed here becauseAbout who the gain goes to: on this claim, anyone who starts now.

1. David Heinemeier Hansson 26 Aug 2026 There's not any accumulation, which is in some ways a great relief.

   Creator of Ruby on Rails, CTO of 37signals

   Their wordsThere's not any accumulation, which is in some ways a great relief.

   spoken · official transcript hear it at 1:24:08 · all korrents from this recording

   ↗DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501youtube.com

Coding agents make every plugin interface you ever blew off trivially usable, which is a large productivity win.

1. Thomas Ptacek 19 Aug 2026, 17:03 UTC I can't say enough how excellent a coding agent is at customizing Chrome with extensions. A huge productivity win. Every plugin interface you've ever blown off is now trivially usable.

   Security engineer; founder of Matasano Security and Latacora

   Their wordsI can't say enough how excellent a coding agent is at customizing Chrome with extensions. A huge productivity win. Every plugin interface you've ever blown off is now trivially usable.

   ↗@tqbf on Xx.com

With a good enough plan, a coding agent implements the whole change in one attempt.

1. Boris Cherny 4 Mar 2026 once there is a good plan, it will one-shot the implementation almost every time.

   Creator of Claude Code at Anthropic

   Their wordsonce there is a good plan, it will one-shot the implementation almost every time.

   spoken · official transcript

   ↗Building Claude Code with Boris Cherny — The Pragmatic Engineernewsletter.pragmaticengineer.com

Most organisations are bottlenecked on ideas, vision and taste rather than on implementation, so making code cheap does not make them productive.

1. David Heinemeier Hansson 26 Aug 2026 The other thing I'd say is that most organizations don't know what they want. They don't know how to make it better. They're not bottlenecked on implementation. They're bottlenecked on ideas. They're bottlenecked on vision. They're bottlenecked on taste. And if you don't have those element in excess of your implementational capacity, it doesn't help. So you can make a lot of shitty ideas come true, then what?

   Creator of Ruby on Rails, CTO of 37signals

   Their wordsThe other thing I'd say is that most organizations don't know what they want. They don't know how to make it better. They're not bottlenecked on implementation. They're bottlenecked on ideas. They're bottlenecked on vision. They're bottlenecked on taste. And if you don't have those element in excess of your implementational capacity, it doesn't help. So you can make a lot of shitty ideas come true, then what?

   spoken · official transcript hear it at 0:20:13 · all korrents from this recording

   ↗DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501youtube.com

The ideas in software originate with humans; agents build what people tell them to build.

Filed here becauseFiled here as the strong form of the same argument — and its holder has since stepped back from it.

1. David Heinemeier Hansson 26 Aug 2026 I thought all the ideas would originate with the humans. They would tell their agents what to build, and off they went.

   Creator of Ruby on Rails, CTO of 37signals

   Recalls holdingI thought all the ideas would originate with the humans. They would tell their agents what to build, and off they went.

   spoken · official transcript hear it at 0:36:24 · all korrents from this recording

   Their words nowI don't think that's true anymore at all.

   ↗DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501youtube.com

Coding agents made implementation cheap without making roadmaps shorter or software better -- the binding constraint was never how fast code gets written.

1. Sunil Pai 14 Aug 2026 buuuut my roadmap is’nt ten times shorter. and I’m definitely not ten times better at deciding what is worth building. teams haven’t started casually shipping a year of product work every month. you can look around and and it’s hard to say what software has gotten meaningfully better in the last year.

   Systems and developer-tools engineer at Cloudflare as of 2026

   Their wordsbuuuut my roadmap is’nt ten times shorter. and I’m definitely not ten times better at deciding what is worth building. teams haven’t started casually shipping a year of product work every month. you can look around and and it’s hard to say what software has gotten meaningfully better in the last year.

   ↗the task isn't the jobsunilpai.dev

A machine can complete the task and completely miss the job -- coding makes this easy to see precisely because its tasks are so legible and verifiable.

1. Sunil Pai 14 Aug 2026 a cybernetic machine can nail the task and completely miss the job. coding makes this very very easy to see because the tasks are so legible (and verifiable!): write the function, migrate the API, fix the tests.

   Systems and developer-tools engineer at Cloudflare as of 2026

   Their wordsa cybernetic machine can nail the task and completely miss the job. coding makes this very very easy to see because the tasks are so legible (and verifiable!): write the function, migrate the API, fix the tests.

   ↗the task isn't the jobsunilpai.dev

A coding agent does not learn from its mistakes the way a person does -- it repeats the same error indefinitely unless a human notices and writes it down.

1. Mario Zechner 25 Mar 2026 An agent has no such learning ability. At least not out of the box. It will continue making the same errors over and over again. Depending on the training data it might also come up with glorious new interpolations of different errors.

   Austrian developer who created the libGDX game framework

   Their wordsAn agent has no such learning ability. At least not out of the box. It will continue making the same errors over and over again. Depending on the training data it might also come up with glorious new interpolations of different errors.

   ↗Thoughts on slowing the fuck downmariozechner.at

The human bottleneck is a safety property: remove it and small, individually harmless coding errors compound at a rate nobody feels until it is too late.

1. Mario Zechner 25 Mar 2026 With an orchestrated army of agents, there is no bottleneck, no human pain. These tiny little harmless booboos suddenly compound at a rate that's unsustainable. You have removed yourself from the loop, so you don't even know that all the innocent booboos have formed a monster of a codebase. You only feel the pain when it's too late.

   Austrian developer who created the libGDX game framework

   Their wordsWith an orchestrated army of agents, there is no bottleneck, no human pain. These tiny little harmless booboos suddenly compound at a rate that's unsustainable. You have removed yourself from the loop, so you don't even know that all the innocent booboos have formed a monster of a codebase. You only feel the pain when it's too late.

   ↗Thoughts on slowing the fuck downmariozechner.at

In an agent-written codebase the agent-written tests are equally untrustworthy; manually using the product is the only reliable measure of whether it works.

1. Mario Zechner 25 Mar 2026 Worse, you realize that the gazillions of unit, snapshot, and e2e tests you had your clankers write are equally untrustworthy. The only thing that's still a reliable measure of "does this work" is manually testing the product.

   Austrian developer who created the libGDX game framework

   Their wordsWorse, you realize that the gazillions of unit, snapshot, and e2e tests you had your clankers write are equally untrustworthy. The only thing that's still a reliable measure of "does this work" is manually testing the product.

   ↗Thoughts on slowing the fuck downmariozechner.at

Running coding agents without approval prompts is a normalization-of-deviance risk: the fact that it has not caused a disaster yet is itself the problem.

1. Simon Willison 31 Dec 2025 I run in YOLO mode all the time, despite being deeply aware of the risks involved. It hasn’t burned me yet... ... and that’s the problem.

   Co-creator of Django and creator of Datasette

   Their wordsI run in YOLO mode all the time, despite being deeply aware of the risks involved. It hasn’t burned me yet... ... and that’s the problem.

   ↗2025: The year in LLMssimonwillison.net

Overly prescriptive instructions make a coding agent worse, for the same reason a micromanaging boss makes a programmer worse.

1. David Heinemeier Hansson 26 Aug 2026 it was actually being damaged by overly prescriptive humans. And any programmer who's had a pointy-haired boss knows exactly what that is like. When the boss walks into the room, doesn't know anything, starts telling you how to program, how to code. What do you do? You sulk. You write shittier code if you're mandated to do things that are against your better judgment. Why would an agent not be the same?

   Creator of Ruby on Rails, CTO of 37signals

   Their wordsit was actually being damaged by overly prescriptive humans. And any programmer who's had a pointy-haired boss knows exactly what that is like. When the boss walks into the room, doesn't know anything, starts telling you how to program, how to code. What do you do? You sulk. You write shittier code if you're mandated to do things that are against your better judgment. Why would an agent not be the same?

   spoken · official transcript hear it at 0:56:38 · all korrents from this recording

   ↗DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501youtube.com

Coding agents should read the shared AGENTS.md convention; insisting on a tool-specific CLAUDE.md creates split-brain problems across a team.

1. Tobias Lütke 25 Aug 2026, 14:34 UTC I’m thinking about banning Claude code at Shopify until they change their mind and read AGENTS.md and .agents/skills etc. Insisting on only reading CLAUDE.md sometimes leads to split brain problems when different team members use different tools. Just unnecessary.

   Co-founder and CEO of Shopify

   Their wordsI’m thinking about banning Claude code at Shopify until they change their mind and read AGENTS.md and .agents/skills etc. Insisting on only reading CLAUDE.md sometimes leads to split brain problems when different team members use different tools. Just unnecessary.

   ↗@tobi on Xx.com

Claude Code is the coding agent to build with.

1. Pieter Levels 24 Aug 2026 · held 8 weeks I won't use SnekAI but I'll probably switch to @xAI soon

   Dutch bootstrapper who builds and runs his internet startups alone

   No longer holdsI won't use SnekAI but I'll probably switch to @xAI soon

   Their words nowYou're treated like a child getting lootboxes just to be able to do your work while you pay massive amounts of money for extra usage tokens

   ↗Claude Code treats you like a child getting lootboxes so I'll probably switch to xAIlevels.io

MCP may turn out to be a one-year wonder, because coding agents that can run arbitrary shell commands make Bash the best general-purpose tool interface.

1. Simon Willison 31 Dec 2025 The reason I think MCP may be a one-year wonder is the stratospheric growth of coding agents. It appears that the best possible tool for any situation is Bash—if your agent can run arbitrary shell commands, it can do anything that can be done by typing commands into a terminal.

   Co-creator of Django and creator of Datasette

   Their wordsThe reason I think MCP may be a one-year wonder is the stratospheric growth of coding agents. It appears that the best possible tool for any situation is Bash—if your agent can run arbitrary shell commands, it can do anything that can be done by typing commands into a terminal.

   ↗2025: The year in LLMssimonwillison.net

MCP servers are worth removing outright: an agent reading the code directly is faster, and pollutes its context less, than one that can reach for a tool unasked.

1. Peter Steinberger 25 Aug 2025 Even removed my last MCP, since Claude sometimes would go off spinning up Playwright unasked when it could simply read the code - which is faster and pollutes the context less.

   Founded PSPDFKit in 2011 and ran it for a decade

   Their wordsEven removed my last MCP, since Claude sometimes would go off spinning up Playwright unasked when it could simply read the code - which is faster and pollutes the context less.

   ↗My Current AI Dev Workflowsteipete.me

Agentic code search has recall that falls as a codebase grows, so the agent does not find the code it needs -- and which search tool you give it barely matters.

1. Mario Zechner 25 Mar 2026 You can give it a Bash tool so it can ripgrep its way through the codebase. You can give it some queryable codebase index, an LSP server, a vector database. In the end it doesn't matter much. The bigger the codebase, the lower the recall. Low recall means that your agent will, in fact, not find all the code it needs to do a good job.

   Austrian developer who created the libGDX game framework

   Their wordsYou can give it a Bash tool so it can ripgrep its way through the codebase. You can give it some queryable codebase index, an LSP server, a vector database. In the end it doesn't matter much. The bigger the codebase, the lower the recall. Low recall means that your agent will, in fact, not find all the code it needs to do a good job.

   ↗Thoughts on slowing the fuck downmariozechner.at

The part of a coding agent that actually acts on code is not AI at all — it is simple systems code wired to ground truth, the way a Makefile is.

1. Thomas Ptacek 2 Jun 2025 The code in an agent that actually “does stuff” with code is not, itself, AI. This should reassure you. It’s surprisingly simple systems code, wired to ground truth about programming in the same way a Makefile is.

   Security engineer; founder of Matasano Security and Latacora

   Their wordsThe code in an agent that actually “does stuff” with code is not, itself, AI. This should reassure you. It’s surprisingly simple systems code, wired to ground truth about programming in the same way a Makefile is.

   ↗My AI Skeptic Friends Are All Nutsfly.io