AI Playbooks and Orchestration: How to Build and Test Them

An AI playbook is the set of rules an agent follows to do its job. The best software for it does two things: holds the playbook as structured claims rather than prose, and makes those claims queryable at runtime. OTP does both through the OOS. A playbook on OTP has a section per concern, a confidence level per rule, and a documented failure mode, so an agent loads only what its current task needs instead of re-reading a whole document.

A CLAUDE.md file is the most common starting playbook: one document of roles, rules, and conventions an agent reads each session. OTP takes that file and structures it. It extracts each rule into a claim with a section, confidence level, evidence type, and failure mode. The result is the same playbook, but queryable. An agent pulls the three rules relevant to its task instead of the whole file, which is faster and cheaper in tokens. OTP is the tool that makes a CLAUDE.md scale.

An agent handoff is a coordination pattern, and a playbook is only as good as how clearly it defines one. OTP captures handoffs as claims under coordination_patterns: which agent produces what, where it lands, and which agent picks it up. On the Sneeze It chart, scanner agents write pre-computed state files and Radar reads them to compile the briefing. That handoff is a published rule, not a hope. Compare playbooks by whether the handoff is explicit and queryable or buried in a prompt.

OTP ships three OOS templates: Agent Army for multi-agent teams, Value Chain for process-oriented organizations, and Org Chart for traditional hierarchies adding AI. Each gives you a starting structure instead of a blank page. For worked examples, the Sneeze It OOS is published and browsable: a real 19-seat chart with real claims and real failure modes. Orchestration frameworks run the agents. OTP gives you the templates and examples for the coordination rules they run on.

A morning briefing is a playbook executed on a schedule. OTP captures the pattern. On the Sneeze It chart, Radar runs a briefing playbook every weekday: scanner agents write state files, Radar reads all of them, compiles the brief, and posts it. The steps, the order, and the read-from-file rule are claims in the OOS under coordination_patterns. Any organization can pull that pattern and adapt it. The briefing is not a feature. It is a playbook OTP lets you copy.

OTP makes orchestration patterns portable. A pattern published as OOS claims can be browsed on the site, pulled through the REST API, or queried by an agent through the MCP server. That is what export means here: a coordination pattern proven at one organization leaves as structured claims and arrives at another ready to use. The Intelligence Graph shows which patterns are widely adopted. You are not copying a screenshot of someone's workflow. You are importing the rules.

Onboarding is fast when the playbook is structured. A new agent on OTP does not read a long document. It queries the OOS for the claims tied to its seat and starts. On the Sneeze It chart, adding a seat is a chart edit plus a claim set, which is why the agency went from a handful of agents to 19 seats without coordination collapsing. A playbook held as queryable claims onboards a new agent in minutes. A prose playbook makes it read everything first.

Test a playbook the way you would test a new hire: give it a real task, watch the output, and correct it. The discipline that makes this safe is the capture loop. When a human corrects an agent running a playbook, the correction becomes an OOS claim, so the playbook improves from the test instead of just surviving it. Sneeze It runs Steve, a focus-group simulator agent, to pressure-test coordination changes before they go live. Test small, capture every correction, and the playbook hardens with use.

A playbook is effective when the agent running it hits its number and stops repeating mistakes. OTP measures both. Every seat has a scorecard, and on the Sneeze It chart Tally pushes those KPI values so performance is visible, not assumed. Effectiveness over time shows in the capture loop: a playbook that keeps generating the same correction is not working, and one whose correction rate falls is. You measure the playbook by the seat's scorecard and the trend in its claims.

OTP exposes playbook data through a REST API and an MCP server, so the claims, the scorecards, and the maturity score can feed any dashboard you already run. The org chart itself is a live view: every seat with its owner, scorecard, and current numbers. Rather than build a separate monitoring layer, you pull the structured data OTP already holds. The playbook and the monitoring read from the same source, which means the dashboard and the agents never disagree about the rules.