Working with agents
How work is split between an AI coding agent and a person, how rules reach the agent, and what we changed after several sessions collided in one repository.
Notes
Several Claude Code sessions, one repo: the file owner table, checked against 13 merges
One commit swept 50 files from another Claude session into a release fix. We built an owner table next; 7 of 13 merges still needed hand merging.
Three agent designs, three judges: the market judge alone would have picked the one that crashed on every KakaoTalk file
Before coding a document viewer, 3 agent designs competed and 3 judges scored them. A crash on every KakaoTalk file and 23 plan gaps surfaced before code.
Commits without a Co-Authored-By trailer: the agent made 26 of the 29 we could check
We checked 29 commits without a Co-Authored-By trailer against agent session logs: the agent made 26, mostly merges. Commands to count your own repository.
Running an AI marketer on a schedule four times a day: the permission table that leaves money, accounts and policy to a person
Our scheduled AI marketing agent reads one permission table every run. An approved ad restart click was still refused by a tool check, so a person clicked it.
84 agents reviewed a release: three refuters per finding, and only the 20 survivors got fixed
SugarSpeller 1.6.025's release review ran 84 agents with three refuters per finding. 20 findings survived, one of them a crash, and the second pass confirmed 0.
"Use hundreds of agents" in practice: 10 workstreams, 59 commits in a day, three running at once
The operator allowed hundreds of agents, but only three code-writing agents ran at once. 4 GB of free RAM and two shared phones set the pace of a 59-commit day.
How to write AGENTS.md: only the rules an AI coding agent gets wrong
AGENTS.md is the file an AI coding agent reads before it works. Ours has 96 lines, lists only the rules the agent gets wrong, and enforces key ones with tests.