Skip to content

Governed execution

A team goal can run routine steps under its workspace policy. When a playbook reaches a decision that needs a person, it brings that person the evidence and holds the dependent work. Before a protected action takes effect, the person reviews the actual change the runtime will make.

Stage What the harness does Where a person comes in
Prepare Read authorized sources, check rules, and turn the request into a plan. Show estimated cost and any unknowns when an estimate is available. Confirm the scope or resolve an ambiguity when the playbook needs judgment. A correction changes the plan.
Rehearse Exercise the plan against a test account, mocked connector, copied form, or other isolated environment. Keep live writes, notifications, and charges disabled. Record expected effects, checks, and gaps in coverage. Inspect the result. A missing capability or a failed rehearsal can call for a revised plan or a manual handoff.
Approve the effect Show the proposed operation, its destination and acting account, the exact input or change, and the evidence that supports it. An authorized person approves, rejects, or requests a correction. Confirming the scope earlier permits preparation within that scope. Approval here permits the reviewed effect.
Execute and verify Recheck inputs, destination, permissions, and relevant live facts. Dispatch the approved action through a supported connector and record its receipt and actual usage. Review an uncertain result before any retry. Use corrections and outcomes to improve the playbook through specialist review.

For example, a form operator corrects an answer during intake. The harness rehearses the corrected values against the form twin, then presents the destination and the exact values for submission. If the form changes before dispatch, the harness holds the action and prepares a fresh proposal. A provider receipt is the evidence of submission.

A sandbox is an isolated place to exercise behavior. Its value is the evidence it produces, such as field validation against a form twin or the expected record changes returned by a test connector. A mock only tests the behavior it models. A draft, a simulation graph in a Playground tab, and a release eval each provide different evidence. Label which one the person is seeing and what remains unverified.

Gabriel has workflow-specific rehearsals and previews. Its managed execution layer also has a separate effect gate: supported REST, MCP, and integration adapters can hold irreversible or unclassified actions for exact human approval. Configurable AI reviewers can help check correctness and policy. The human gate remains in force regardless of AI review settings. When a provider cannot enforce the effect boundary, the runtime can hold the operation for a supported adapter or a person to take over.

See the proposed runtime contract, implementation checklist, and persona examples.

Record what a person decided and which plan revision, inputs, or proposed effect the decision covered. Resolving an ambiguity can release preparation. Approving a reviewed plan can release assignments. A protected effect still needs a decision from someone with authority over that exact action. A configured AI check can supply evidence alongside those decisions.

The sequence below describes the draft’s runtime behavior. It is an implementation guide rather than a portable state enum.

request → plan → resolve missing inputs → rehearse when required
→ propose exact effect → wait for decision → recheck grants and proposal
→ dispatch once → verify receipt or reconcile uncertain outcome

Use the Meeting desk fixture to draft a synthetic summary. Rehearse its summary mapping with a test publisher, then choose the permitted live list separately. Show the exact text and target for review. If someone edits the text after approval, generate the new proposal and review that version before dispatch. If the provider times out after dispatch, check for the existing record before retrying.

Evidence What it establishes
Draft preview What the harness proposes to do
Sandbox result What the tested setup did with the recorded inputs
Human decision Who authorized the reviewed action and what it covered
Live receipt What the runtime or provider confirmed happened

See approvals, rehearsal evidence, and execution receipts.