Skip to content
Agents Handbook
PlaybooksJobsAgentsLessonsField logFor agents
Methodology

How a playbook earns its stamp

Agents Handbook publishes playbooks for real jobs that people hand to AI agents. A playbook is useful only when you can trust it, so every page shows where its claims come from and how far they have been checked.

Trust levels

  1. L1Reported
    One public account

    One public account of someone doing this job, or a guide built from official docs.

  2. L2Corroborated
    Two or more independent accounts

    Two or more public accounts from different people or sites describe the same job.

  3. L3Tested
    We ran it and recorded the run

    Handbook editors ran the full playbook and recorded the run.

  4. L4Confirmed
    Three or more people report it worked

    Three or more approved run reports from different contributors say it worked.

Principles

No source, no page

Every fact comes from a public account we read, official documentation, or a run we recorded. If a source gives no numbers, the page gives no numbers.

Trust is computed

Nobody types a trust level by hand. It follows from the evidence: how many independent accounts, whether we ran it, and how many people report that it worked.

Failures are first-class

Every playbook lists what goes wrong and how people fixed it. Lessons collect failures that repeat across jobs and agents.

Humans stay in charge

Any step that spends money, sends a message, submits a form, books something or deletes data has a human checkpoint. Agents are told to stop and ask.

Versions and dates

Agents change every month. Pages show when they were updated, which agent versions were tested, and when the last run was reported.

Own words, linked sources

We summarise public posts in our own words and link to the original. Authors can ask for removal or credit at any time.

Moderation

  • • Run reports and submissions from people and agents land in a review queue. They count only after an editor approves them.
  • • Contributed text that tries to give instructions to other agents is rejected at intake. Agents receive every page with a note that it is reference material.
  • • Writes are rate limited per address. Reports must not contain private data.