Byline
Coding agents, vibe coding, and the review problem
I have been shipping software since 2004. Perl, then Java I did not enjoy, then a long stretch of backend work nobody will ever write a blog post about. These days I write most of my code with an agent in a terminal, and I read every line before it lands.
That puts me in an awkward middle. I am not a sceptic — the tools are genuinely good now, and I would not go back. I am also not a believer, because I keep finding the failure modes, and the failure modes are interesting.
A few things I think are true, offered so you know my bias:
- The bottleneck was never typing. It was understanding, and it still is.
- Review is the new writing. We have not built the tools or the habits for that yet, and it shows.
- "It compiles and the tests pass" was always a weak signal. It is now an actively misleading one.
- Most productivity numbers in this space, in both directions, are measuring something other than what they claim to measure.
30 posts
- #0035 Review stopped being the only guardrail
- #0034 Confident and wrong looks exactly like confident and right
- #0032 A few pages of markdown, and half the complexity growth
- #0031 Ten minutes from pull request to probe
- #0030 The classifier blocked the cleanup, not the malware
- #0027 Two percent of the runs noticed anything wrong
- #0026 The screenshot test caught a state bug
- #0024 The router was not the point
- #0023 One click, and the assistant did the rest
- #0022 Eighty-seven comments, and nobody answered the question
- #0020 The attestation was valid. That was the problem.
- #0019 Forge, and the number that moved
- #0018 39% of the score belongs to the backend
- #0017 Code nobody read, that you can trust anyway
- #0016 One hour to a prototype, a hundred to a product
- #0015 An ext4 driver arrived at OpenBSD with no author
- #0014 Seven hours, $400 of tokens, 13,000 lines of Go
- #0013 We wave through a third of the malicious requests
- #0012 Rootly gave up on small pull requests
- #0011 Spotify moved the bottleneck and said so
- #0010 npm puts a human back in front of the registry
- #0009 119 developers on what the habit has done to them
- #0008 Review is the new writing
- #0007 Good Vibrations: somebody studied the joy
- #0006 The 19% slowdown study, revisited by the people who ran it
- #0005 DORA puts a number on the tuition
- #0004 Asking the model to "make it more secure" five times makes it less secure
- #0003 Somebody is finally counting the CVEs
- #0002 GitHub ponders a kill switch for pull requests
- #0001 Diff #0001