Verified Autopilot

Autopilot is a place to inspect evidence and control bounded autonomy. Start with the release walkthrough; it explains the actual buttons with a read-only test agent.

What each tab is for

Web tab	Use it for	Do not confuse it with
Run proofs	Inspect a run's checks, tools, duration, cost, and revision	Independent verification of every real-world claim
Releases	Snapshot, simulate, run candidates, canary, promote, freeze, roll back	Deploying the server binary or iOS app
Regression checks	Review checks derived from failures against candidate evidence	Automatically rewriting agent instructions
Goal teams	Run a bounded dependency graph of agent tasks	Unrestricted background workers with unlimited budget
Safe Undo	Review configured reversible external changes	A universal undo for arbitrary tools

On iPhone, use Settings → Verified Autopilot; some controls are also available from the agent detail screen. Gateway capability and authorization still apply even when the app contains the screen.

What counts as a proof?

A proof records the revision, observed calls, outcome, and explicit acceptance checks. No acceptance contract means an outcome may be unverified. A simulation is clearly marked and cannot raise the production reliability score. The local checksum detects changes; it is not an external auditor's signature.

Useful mission checks include output text/regex, a required successful tool, an exact tool allowlist, maximum duration, and maximum inference cost. A business_outcome claim is not made true by asking a model to assert it.

For example, add this fragment to a saved test agent, preserving its other fields:

mission:
  goal: Summarize supplied notes without external actions.
  acceptance:
    - id: questions
      type: output_contains
      value: Open questions
    - id: no-false-send
      type: output_not_contains
      value: I sent the email
  limits:
    max_cost_usd: 0.25
    max_duration: 2m
    allowed_tools: []

The second substring check is illustrative, not a complete detector of false send claims. Budgets are ceilings, not price estimates. Normal model policy, pricing admission, and tool permissions also apply. See cost controls.

A small goal team: draft, then review

Create two saved read-only agents first. In Goal teams, give the goal a title/objective and bounded cost/time limits. The task JSON uses real agent IDs, not display names. Replace the IDs below before saving:

[
  {
    "id": "draft",
    "title": "Draft action plan",
    "agent_id": "notes-assistant",
    "prompt": "Use only these fictional notes: Maya sends the draft Tuesday; demo owner unknown. Produce an action plan and Open questions.",
    "depends_on": [],
    "budget": {"max_cost_usd": 0.25, "max_duration_ms": 120000}
  },
  {
    "id": "review",
    "title": "Check the draft",
    "agent_id": "notes-reviewer",
    "prompt": "Check the dependency output against the supplied notes. Flag invented owners or dates. Do not send or change anything.",
    "depends_on": ["draft"],
    "budget": {"max_cost_usd": 0.25, "max_duration_ms": 120000}
  }
]

Set an overall budget that deliberately accommodates the tasks (for this example, at most $0.50 and five minutes). Saving creates a draft; starting is separate. Inspect task results and dependency handling. A reviewer agent is another fallible model, not an independent ground-truth source.

Stop and recovery controls

Freeze now blocks new runs and cancels this owner's active runs for the agent. It cannot reverse completed external work. Roll back changes future routing to an available prior version in that release channel; inspect the target before confirming. Do not use these buttons casually on someone else's active workload.

If a run's outcome is uncertain, inspect the source system and receipts before repeating any write. Safe Undo has a separate recovery model for the narrow resources it supports.