A playbook is a recurring job your AI workers run — but only when there's evidence it should. Instead of an agent waking on a schedule and deciding what to do, a playbook wakes on its triggers, checks its evidence rules against your real data, and either acts or records a quiet inspection and goes back to sleep.

That inversion is the point: no evidence, no work. A playbook can never invent tasks to look busy.

Anatomy of a Playbook

Part	What it does
Goal	The business outcome, in plain language. Scopes what the worker is allowed to pursue.
Triggers	When to wake up: a schedule, record events, or both.
Evidence rules	Whether to act: deterministic checks against your data. All rules must pass.
Worker	The AI agent that does the work when evidence passes.

Find them at Playbooks in the app. Any focus area becomes a playbook once you give it a trigger and a worker.

Triggers: when to look

Triggers are deliberately coarse — they decide when to check, not whether to act.

Schedule — pick from presets (hourly, every 6 hours, daily, weekday afternoons, weekly, monthly) or write a custom cron expression, with a timezone.

Event triggers — wake when records are created or updated in a module you choose, checked about once a minute. You can add up to five, mixing types and modules.

Events OR each other: any one trigger with fresh signal wakes the playbook.
Evidence rules AND each other: all rules must still pass before it acts.
A burst of activity produces one run, not one per record — and anything left unhandled wakes it again on the next check.

Events only — choose "No schedule (events only)" to make a playbook purely reactive. It never runs on a clock and stays completely silent until something actually changes.

You can also combine both: react to events as they happen, with a daily sweep as a backstop.

Evidence rules: whether to act

When a playbook wakes, its rules are evaluated as plain database checks — no AI model is invoked to decide whether work exists.

Rule	Passes when…
New records since last run	At least N records were created in a module since the playbook last acted
Records gone stale	At least N records haven't been touched in X days
Records matching a filter	At least N records currently match your criteria
Cooldown since last action	The playbook hasn't acted within the last X hours
Quiet hours	The current time is outside a window you define

Field filters

The record-based rules accept up to five field filters, so evidence can be as specific as your data:

New records in Tasks where Notes contains "Auto-filed from Sentry"

Filters support contains, doesn't contain, and equals on the record's Name, Notes, or any custom field. Records that don't match simply don't count as evidence.

Quiet runs are healthy

When rules don't pass, the playbook records a quiet run: which rules failed and why, zero tasks created, zero AI cost. The run ledger on each playbook shows every wake-up — acted, quiet, or blocked — so "the agents did nothing this week" is something you can audit, not something you have to trust.

Run now is always available and goes through the same evidence gate — a manual poke can force an inspection, never invented work.

What happens when a playbook acts

Exactly one task is created for the bound worker, carrying the evidence that justified the run and strict scope discipline: work only this evidence toward the goal, don't invent side projects, and finish honestly if the evidence turns out to be insufficient. The task then flows through everything you already control — approval rules, autonomy levels, and quality review.

Converting legacy agent schedules

If an agent has its own schedule from before playbooks, the Playbooks page shows it under Legacy agent schedules with a one-click convert: the schedule moves onto a new playbook, the agent becomes its worker, and the old schedule is disabled (re-enable it to roll back).

Converted playbooks start with no evidence rules — they'll act on every firing until you add some. Add rules right after converting; that's the step that ends work-for-work's-sake.