Skip to main content
The Agent Guide tab in a job’s workspace shows what every agent managing that job receives, how current it is, and what needs your attention. It is also where your team writes and publishes the Playbook, the human guidance that agents follow underneath the job’s typed rules. You do not need to set anything up for agents. When a teammate’s agent manages the job through the OpenTrain CLI, MCP server, or SDK, it receives the Agent Guide automatically and is told exactly what changed after each publication.
If you do not see the Agent Guide tab on a job, it may not be enabled for your organization yet. Email support@opentrain.ai to ask about availability.

​
Open the workspace

  1. From My Jobs, open the job.
  2. Select the Agent Guide tab.
The workspace has a navigator with seven sections and a detail panel. On narrow screens it opens as a list first; use All sections to return to the navigator. Canonical facts always outrank the Playbook. The Agent Guide section shows the authority order so everyone reads it the same way: platform rules, typed policies and workflow, recorded decisions and exceptions, then the published Playbook, then team, personal, and session notes.

​
Write and publish the Playbook

Open Playbook, then Draft & publish.
  • One shared draft per job. Everyone with access to the job edits the same draft, and your changes save automatically. If a teammate saved first, the editor tells you and lets you reload their draft instead of overwriting it.
  • Headings define sections. Each heading starts a section with a stable id, so agents can follow exact changes to that section across versions. Keep paragraphs, lists, code, and quotes under a heading.
  • Publish when ready. Click Publish, add an optional publication note, and confirm. Publishing creates a new immutable version, records who published it, and tells agents exactly which sections were added, changed, or removed. It never edits an earlier version.
  • Only job managers publish. The organization owner, or a team member assigned to the job, can publish. Agents cannot publish; they propose.
The draft is stored for the editor only. Agents receive the published version as Markdown sections with stable ids and content hashes, never the editor document. If someone publishes while you are editing, the editor asks you to reload from the current version before you can publish again. Nothing you typed is lost until you choose what to do.

​
Review agent proposals

Open Playbook, then Agent proposals. Agents can propose a new or changed section, propose removing a section, or leave a note. A proposal is a recommendation only. Nothing an agent proposes is followed until a human publishes it. For each open proposal you can:
  • Apply to draft. Insert the proposed section into the shared draft exactly as proposed, so you can review it in context before publishing.
  • Accept with the next publication. Select the proposal so it is marked accepted when you publish a version that contains exactly the proposed content. OpenTrain checks the content before anything is written; if the published section differs, the publication is refused and nothing changes.
  • Decline. Close the proposal with an optional note. The agent that made it can see the outcome.
Proposals written against an older version are flagged as stale so you can decide whether they still apply.

​
Track versions and agent activity

  • History lists every published version with its publisher, publication note, content hash, and the sections added, changed, or removed. View shows a version exactly as agents receive it, and Changes shows the exact delta between two versions.
  • Activity shows recent shared-context events and, for each teammate’s agent session, the last shared-context revision it received and acknowledged, so you can tell which agent is current and whose agent synced.

​
Job Instructions are separate

The Playbook guides agents and managers. It does not change what AI trainers see. Worker-facing guidelines stay in the job’s Instructions tab, which remains the only canonical source of instructions for hired AI trainers.