🤖AI coding

Monitoring the situation

A good way to stay abreast of AI tooling: Rands Leadership Slack #ai-coding-tools

Document driven (need link)
An iterative approach that makes your genie less insane over time Stdlib from Ghuntley (Cursor). No longer available to nonsubscribers, but the basic idea is to pay attention to your workflow each time, instead of just banging your head against the wall.
A structured approach to coding by Ryan Carson. For Cursor but generally relevant.
Kent Beck described a TDD and tidy first approach he calls augmented coding.
Vibe engineering describes patterns for those who are building high quality code quickly with AI.
Why AI swarms can’t build architecture and what to do about that.
Consider using ast-grep for larger codebases (and also this tool looks amazing).
Potential tooling for reducing duplication in code: jscpd, pylint for Python, dupl for Go.

Humans ON the loop by Kief Morris. Articulates a lot of the direction my thinking has been going around where a lot of engineering focus will probably be. Must read.
Pairs nicely with Lethain’s What can agents do which describes some of the nuts and bolts of what it takes to have an internal agent you have to support.
Stripe’s Minions show a workflow for creating pull requests in a large codebase, triggered by Slack. Part 2 goes into some detail of how it works. Most interesting part is the orchestration and how it separates deterministic and LLM based approaches. It’s a mostly one-shot approach, and suffers a bit because of that. But very interesting!
Harness engineering for coding agent users. Describes some useful concepts for understanding how to engineer more reliably with coding agents.
Dark factories this post is mostly alright but the point I found useful is the concept of “dark factories”, factories where you are not designing for people but machines.
Harness engineering at OpenAI A Feb 2026 article on an internal team at OpenAI that instituted a “no human code rule” and went from 0 -> 1MM lines of code. Contains some useful things they learned.
Digital twin universes approach to cloning dependencies for fast automated verification.
Intercom’s Fin on how they 2xed PRs, contains a lot of good metrics and info on how they went about it.

How to kill code review from Ankit Jain, founder of Aviator. Argues code review isn’t sustainable any more, and we have to start developing approaches like BDD and deterministic verification.
Intercom on AI approval of PRs. 19% of PRs approved by AI. Note I think their metrics are bogus, since of course the easiest and safest PRs will be reviewed by AI. Some nice approaches to how they did it, however.
Ratchets and linting as a tool for increasing the amount of quality in automated codebases.
PR review tips from posthog.
Automating PR review at Fin/intercom

The Gap Through Which We Praise the Machine on usability issues in agent coding, and why some people seem to be doing well and others aren’t.
Nobody codes here anymore A case study on adoption

Building eval systems that improve your product described a methodology for creating evals. Start with error analysis and an expert, build eval suite of code and LLM evals. Pass/fail based, with critique. Split data into train, dev, and test sets. True pass rate over true negative rates. Paywalled but useful.

Revenge of the Junior Developer - Steve Yegge (who has a vested interest in this) lays out his vision of where this is going: agents, agent fleets, etc. March 2025.

Should probably have its own page

Charity Majors: make AI boring again argues we should engage in using AI especially if we have concerns about it.