When an agent finishes real work, a second agent reads how it went, starting from three questions. What it finds goes into the instructions the next agent starts from.
As of September 2026 this is a standing rule in my repo, not something I try to remember. When one of the agents that writes code finishes a substantial run, the agent coordinating the work is told, in my repo's AGENTS.md, to start a retro agent, usually in the background.
The retro reviews how the work was done, not the code. Another agent reviews the code.
My repo's rules name these three. The agent's own definition adds three more: dead ends, whether it was the right approach, and whether anything is worth writing up. What happens to the scripts it produces is in The third time is one command.
It fixes the safe things itself: a line in an agent's definition, a convention, a small script. Anything that needs me, like a new dependency, a production change or spending money, goes at the top of its report as a decision. It does not guess.
A hook records every finished agent run in a queue file. It starts nothing. It only means a retro I skipped is still on the list.
Once a pattern has run clean three times in a row, with nothing for me to tidy up, repeat runs get a retro only when they did something new. Otherwise the retro keeps reporting "nothing to learn", and that is busywork.
I have not counted how often a retro finds nothing, or what it costs against what it saves. And it only helps where the next agent starts. Once, a retro's fixes sat on a feature branch, and agents started meanwhile read the old instructions.
After any substantial agent run, run a retro on the process, not the code. Ask: could it have been done more directly? Ask: was there repeated manual work that should become a script or command? Ask: did it hit a gotcha or wrong default that belongs in the agent's instructions? Apply the safe fixes now. Put anything that needs a human decision at the top of the report.
When an agent finishes real work, a second agent reads how it went, starting from three questions. What it finds goes into the instructions the next agent starts from.