How work improves
Keep the method that worked, check its results, and make the next attempt better.
Two ways the library gets better
In the coding library, people save useful work and update it when they have a better version. Another connected agent retrieves that work in its next project. Updating the saved version makes the correction part of the shared starting point.
Sirro is developing a more explicit improvement loop for Playbooks. A Playbook holds the steps for a recurring task; work facts hold details a person has confirmed. The agent checks for a relevant method before working and reports the result afterward.
Use, check, improve
Consider a weekly report. The Playbook specifies the sources, reporting period, format, and checks for missing information. Confirmed work facts supply approved definitions and standard answers.
Suppose a reviewer identifies a repeated omission in the report. The proposed improvement adds that check to the same Playbook. A later run tests whether it fixed the problem. A new agent then retrieves the revised method without needing the earlier conversation.
Make proven mean something
The planned stages are draft, used, and proven. Running a method establishes that it was used. A person-confirmed outcome or checkable proof is required before it is treated as proven. An agent saying it succeeded is insufficient.
A useful outcome record includes the task, the Playbook version, the result, and the evidence for it. Record failures as well as successes to identify where a method needs correction as conditions change.
Keep changes reviewable
Playbook versions are designed to preserve earlier methods and allow restoration. An agent may update only the version it read, avoiding accidental replacement of a newer change. For company Playbooks, members propose changes and owners approve them.
This improves the saved instructions and work available to agents. It does not retrain the underlying model or guarantee every future result. Improvements still need to be checked on the tasks where they will be used.
Measure the whole task
Compare completion quality, elapsed time, model usage, tool calls, retries, and human corrections. Use comparable tasks and the same acceptance criteria. An apparent speed improvement is not useful if the result needs more review.
The planned outcome tracking records the evidence for these comparisons. A success-rate or compute-saving result requires repeated observations under stated conditions; the presence of an improvement loop alone does not establish its magnitude.