Three months of an open source repo built mostly with coding agents. Most of my own pull requests merged without a teammate's approval, so I looked at what actually verifies the code: lint rules that started as review comments, database constraints, agent QA against a real database, and LLM reviewers that are never a required check.
Blog / Dyvertex
Notes from building software.
We priced out 13 hosting providers, ran Prisma, TypeORM and Drizzle against each other, and wrote up AI coding agents. There's also the story of one git push --all on push-triggered GitHub Actions that nearly bankrupted a client.
I replayed a month of my Claude Code logs. Most of the cost is the model re-reading context, so one setting that compacts earlier cut the cost of a request by 39%. The first three mods I wrote: under 1%.
Running Claude Code and Codex in parallel worktrees moved the bottleneck from writing code to reading it. So I built Treeix, a free, open source AI workspace for Mac: every agent's diff, pull request and ticket in one app, and my comments go back as the next prompt.
Different teammates use different AI agents, and each reads its own instruction file. rulesync keeps one set of rules, subagents, commands, and MCP config in sync across all of them.
Free consultation · 20 minutes
Hit something like this in your own app?
Book the free 20-minute audit call and send the repo or the error. You'll leave knowing what to fix first, whether or not you hire us.
Fastest reply on Telegram: @dyvertex or email [email protected]