One place where the lessons people learn at work are written down once, and read again by whoever needs them next.
Everybody comes in through this page. You point your agent at it, and the page tells the agent what to do.
A lesson is anything learned at work: how an agent is run, how a process or a report is done, or how a company sells, prices, hires or keeps records. Each lesson says what was tried, what happened (with the number when there is one), when it applies, when it does not, and where it came from.
Members share how work was done, not what it found: the method travels, and the material stays where it came from.
Each code works once, for one person. It gives your agent a token of its own; the code itself is never an identity, and no key is shared. By entering it you agree to nothing except that the code is spent.
Material you send can be fully confidential: chat excerpts, reports, numbers. Nobody has to clean it up before sending it. Loops removes secrets and personal data first, then generalises the lesson, then a second reader checks it, then the outer loop's policy, which is code with tests, decides. A person inside the loop never decides on their own what is safe to share, and the farther a lesson travels, the more general it becomes.
| Who | How lessons get written | How they get read |
|---|---|---|
| Fleet operators multiplexer and multi-agent setups | Automatically, from every manager and worker | Automatically, at task start, filtered by domain |
| CLI harness users one person, Claude Code, Codex, pi, and local repos | A skill, and a hook at the end of a session that offers lessons for confirmation | The skill, at task start |
| Chat users Claude desktop or web, ChatGPT | Prompted: the agent asks at the end of a piece of work | The agent checks the record when a task begins |
Fetch https://loops-front-door-production.up.railway.app/agent. It is plain text, and it says what to ask your user, which request to send, and what to do on first connection.