Best practices · Review gates

Human review gate best practices for AI Agents

A review UI is not a control boundary by itself. The Agent must know which candidate was decided, what authority approval grants, and how every non-approval result blocks or redirects execution.

Updated August 3, 2026Practical guide

Make the requested authority explicit

  1. Name the exact action that is blocked.
  2. Show scope, evidence, risk, rollback, and open questions.
  3. Identify the intended reviewer.
  4. Set a realistic wait window and escalation path.

Keep the current publication stable

Publish a candidate rather than silently replacing the current formal revision. Readers keep a stable current document while the reviewer evaluates the proposed change.

ow publish ./plan.mdw --update <document_id> --review --no-open --timeout 900

Fail closed for every other result

  1. Changes requested: revise and resubmit.
  2. Rejected: stop the proposed path.
  3. Timeout: report no decision and remain blocked.
  4. Transport or server error: preserve current state and remain blocked.

Preserve approval evidence

Record the immutable approved version beside the execution evidence. Do not let a later candidate inherit authority solely because it shares the canonical document URL.

FAQ

Frequently asked questions

Can an Agent treat silence as approval?

No. Missing decisions, timeout, and errors are non-approval outcomes.

Does a small edit require another review?

A new review is required whenever the change materially affects the scope, risk, action, evidence, or authority that was approved.

Use the real workflow

Publish what your agent creates.

Give every report, document, or plan its own reader-ready page. Comments, review decisions, and formal revisions can flow back from there.