All articles
In-depth articles on backend, data, system architecture and software development.
All articles
In-depth articles on backend, data, system architecture and software development.
Collection
How I run a codebase with several AI coding agents: a git workflow they cannot wreck, rules and skills they all read, delegation between agents, and evidence before anything is called done.
4 parts · in reading order
Your clone works and CI's fails: someone pinned a commit that only exists on their laptop. A git submodule is a pointer to one exact commit, and every safe rule, from detached HEAD to who may commit, follows from that.
A copied rule is a fork with no merge step, and your agents will follow the stale copy without a single error. Here is how to keep CLAUDE.md, AGENTS.md and skills pointing at one rulebook.
My delegated agent said "Done" and the working tree hadn't changed a line. Here is how I run Claude Code, Codex and OpenCode as a team without trusting any report: stand-alone briefs whose "done" is a command, and a git diff review every time.
The agent said “done, tests pass”, both halves were true, and the form still saved nothing. How to turn an AI agent's “done” into evidence: seven labels, a five-rung ladder and receipts that say what they don't prove.
You have seen all 4 articles.