AgentsRoom Learning Loop: turn the corrections you keep repeating into a rule every AI agent follows
Filmed in the real app: the same corrections typed again and again to two Claude Code agents and a Codex agent (run the tests before you commit, leave the lockfile alone). The graduation cap in the dock opens Learning Loop, Run the review, and an agent reads those prompts with the user's own CLI. Each proposed rule shows the sentences typed and how often; one click on Accept writes it into the project memory. A brand new Codex agent, whose task ends with a commit and says nothing about tests, reads the rule, runs the tests on its own and commits once all 15 pass.
Read the full feature pageTranscript
Correct your AI agents once, and every agent remembers. This is AgentsRoom, the desktop app where your AI coding agents work side by side: two on Claude Code, one on Codex. All morning, I kept typing the same thing: run the tests before you commit.
To the Codex agent too. And leave the lockfile alone. Each session starts from zero.
In AgentsRoom, the graduation cap in the dock opens Learning Loop. I click Run the review. An agent reads what I typed to this project's agents, with my own CLI.
Then the rules come back. Each rule shows the sentences I typed, and how often I said it. I accept the one I want.
Nothing reaches my agents before that. It goes into the project memory, read by every agent, whatever its CLI. Now I start a brand new agent, on Codex.
Its task ends with a commit. Not a word about tests. Before working, it reads the project memory, and the rule is there.
Then, before committing, it runs the tests on its own. They catch a change, it fixes it, and runs them again. Only then does it commit.
All fifteen tests pass. Correct once, every agent remembers. Learning Loop is in AgentsRoom, and the link is in the description.