Tejas Kumar

Code Is Cheap: How Agentic Coding Changed Collaboration

This post is about how working with other engineers changed for me once agents started writing the code. Code is cheap now: Claude writes mine so I’m not attached to it anymore, which means a teammate pushing a fix into my pull request feels like help instead of an insult. With no code of my own to defend, my team and I spend that attention on the product and on the people who use it.

I was in San Francisco last week for the AI Engineer Code Summit and this post came out of a conversation I had there with Sunil Pai and Addy Osmani. They’re both members of technical staff at Anthropic: Sunil joined this month and Addy works on Claude Code after more than 14 years at Google. Thank you both!!

I used to hate opening pull requests

When I was starting out as a junior engineer I really valued my code (I’d spent a lot of time writing it). Then I’d open a pull request and the comments would come in.

Collaborating was fine, but sometimes the comments were condescending and I’d get annoyed. Other times they were kind of useless, like ā€œhey, let’s rename folder to directory in the codeā€ and I’d think: bro, this doesn’t matter. Sometimes my PR would just get rejected and I’d take it personally.

Maybe the worst was when somebody pushed their own commits to my branch, on my PR, without talking to me first. I took that so offensively. Who do you think you are?!

I wasn’t the only one. Researchers at Google asked 1,317 of its developers how code review felt and 57% of them said they have a negative experience with it at least once a quarter (26% said at least once a month). An earlier study of code review at Google listed tone and power among the ways a review breaks down, where power means ā€œusing the code review process to induce another person to change their behavior; for example, dragging out reviews or withholding approvals.ā€

Why it felt so personal

There’s a 2012 paper called ā€œThe IKEA effect: When labor leads to loveā€ where people either built a plain IKEA storage box or were handed one that was already built. Then everyone bid on their box.

Average bid for the same IKEA storage box (Norton, Mochon and Ariely, 2012)
People who built it
$0.78
People who didn't
$0.48

The builders bid 63% more for the same box!! My code was my IKEA box. I’d put the hours in so it was worth more to me than it was to the person reviewing it. And every comment sounded like somebody telling me my box was only worth $0.48.

Gerald Weinberg called the fix egoless programming in his 1971 book The Psychology of Computer Programming: review each other’s work as friends and keep your feelings out of it. That’s easy to agree with. It was a lot harder to do with my own work on the screen.

What does ā€œcode is cheapā€ mean?

ā€œCode is cheapā€ means that an AI coding agent can write working code in minutes. So the hours a person used to spend typing it are no longer what makes software expensive. Deciding what to build, checking that it works and answering for it once it ships cost as much as they ever did.

Simon Willison says it like this in Writing code is cheap now: ā€œDelivering new code has dropped in price to almost free… but delivering good code remains significantly more expensive than that.ā€ Sunil wrote the same thing from the other side in the context is the work back in January: ā€œagents make code cheaper; they don’t seem to make judgment cheap.ā€

Then Claude started writing my code

Now it’s different. I don’t value my code that much because Claude wrote it. I didn’t build the box. :)

So when something isn’t right in a pull request I opened and a teammate co-authors a fix with Claude and says ā€œhey, I’m just going to fix this and push it into your branchā€, I’m like: this is fine. More people collaborate in my PRs now than before and I welcome it. We’re all just using Claude (or some agent) anyway.

The exact thing I used to take so offensively is something I’m glad about today. On GitHub you can even see it happening in the commit message, where a Co-authored-by line credits everyone who worked on a commit. This is the one Claude Code adds:

Fix the empty state on the settings page

Co-Authored-By: Claude <noreply@anthropic.com>

When 2 people and 2 agents have all touched a branch, ā€œmy codeā€ stops meaning much. Sunil has a line for what’s left to look at: ā€œthe implementation ā€˜matters’ less; you’re reviewing an interpretation, and the history that brought you there.ā€

What happens to code ownership when agents write the code?

When agents write the code, the team that ships it owns it instead of the person who typed it. Martin Fowler described this in 2006 as collective code ownership: ā€œThe code base is owned by the entire team and anyone may make changes anywhere.ā€

Junior me would never have signed up for that (my branch was my branch). Agents got me there without anybody asking. Nobody spent days writing the lines so nobody guards them.

I still answer for what my PR does though. Addy says so in Agentic Code Review: ā€œA model cannot be paged and cannot be held responsible for what it shipped, so whoever clicks merge owns it.ā€ Armin Ronacher, who wrote that an infrastructure project at his new company is more than 90% AI-written, is just as blunt: ā€œI still treat every line as my responsibility, judged as if I wrote it myself. AI doesn’t change that.ā€

Cheap code still needs review

Here’s the part where I argue with myself. Not being attached to code can slide into not reading it. The numbers on that are rough!!

Faros AI looked at more than 10,000 developers on 1,255 teams for its AI Productivity Paradox report in 2025. The teams using AI the most merged way more pull requests. Those pull requests were also bigger and they waited longer for a reviewer.

Change on teams with high AI adoption (Faros AI, AI Productivity Paradox Report, 2025)
Pull request size
+154%
Pull requests merged
+98%
Time in review
+91%

A year later Faros measured 22,000 developers and found that the number of pull requests merged with no review at all had risen 76.3%, up from a 31.3% rise 6 months earlier. Addy describes how it happens: ā€œReviewers simply could not keep pace with the volume, so code began merging unread, and that became normal.ā€

Anthropic ran into this as well. ā€œCode output per Anthropic engineer has grown 200% in the last year. Code review has become a bottleneckā€ is how it announced code review in Claude Code in March. Now Claude looks at nearly every pull request there first and a person still approves it.

Share of Anthropic's pull requests that get substantive review comments, before and after Claude reviewed them first (Anthropic, March 2026)
BeforeAfter
Reviewed16%54%

I’ve felt the cost of not reading myself. When I let a software factory build one of my products for days, I ended up asking my own project what it had done, which is called cognitive debt and it’s no fun.

How to review AI-generated code

If your team is where I was, this is where I’d start:

  1. Run it first. Look at what the change does for the person using the product before you read how it’s written. Simon’s version is that your job is to deliver code you have proven to work.
  2. Make sure the author can explain it. Addy’s rule in his post on reviewing AI code is ā€œNever commit code you can’t explain.ā€ It goes for the person who prompted it too.
  3. Read what was asked for. The request the agent got tells you more than the diff does. That’s Sunil’s ā€œinterpretationā€ again.
  4. Push the fix. If folder really should be directory, have your agent rename it and push the commit. It’s quicker than writing the comment and nobody has to feel anything about it.
  5. Say what you pushed. One line in the PR is enough so nobody finds commits they didn’t expect.

If your team is figuring out how to work with agents, I run workshops.

Code was always a means to an end

This is what I’m seeing across the board with agentic coding: we tend to focus more on the product. When nobody is defending their code, the conversation in a pull request is about whether the thing works and whether it’s the right thing to build.

And it’s so wonderful because code was only ever a means to an end. We get to spend more of our time thinking about the end, which is such a blessing. In fact, the end isn’t even the product: it’s what we’re putting in front of the people on the other side of the screen.

I’m so here for this shift.

Questions

What does "code is cheap" mean?

"Code is cheap" means that an AI coding agent like Claude Code can write working code in minutes. So the hours a person used to spend typing it are no longer what makes software expensive, while choosing what to build and proving it works stay as expensive as before.

How do I review AI-generated code?

Review AI-generated code by running it and checking what it does for the person using the product before you read how it's written. Make sure the author can explain the change, read the request the agent was given next to the diff, and when a fix is small, push it to the branch instead of leaving a comment.

What happens to code ownership when AI agents write the code?

When AI agents write the code, the team that ships it owns it instead of the person who typed it. Nobody spent days writing the lines so nobody guards them, which is what Martin Fowler called collective code ownership in 2006. Whoever merges a change still answers for it.

Is it OK to push commits to someone else's pull request?

It depends on the team. When I wrote my code by hand I took it as an insult. Now that an agent writes my code I welcome it. A quick message saying what you fixed is still polite. GitHub records who did what with a Co-authored-by line in the commit.

Does AI make code review less important?

No. AI makes code review more important and harder to keep up with. Faros AI measured 98% more pull requests merged and 91% longer review times on teams that use AI heavily, and Anthropic says review became its bottleneck after code output per engineer grew 200% in a year.