Expected outcome
Your team has a written review boundary that names the exact pause, the evidence a reviewer needs, the permitted decisions, and the safe fallback.
You can use this public guide without a Tempera account, product credential, or private workspace. Product access begins only after private onboarding.
Before you start
- One bounded browser-agent workflow
- The outcome the agent is expected to produce
- A person who understands the cost of a wrong action
Steps
- 01
Start with the consequence
Identify the action whose cost, irreversibility, or ambiguity justifies interruption. A low-risk, reversible step usually needs observation rather than approval; a refund, external message, permission change, or destructive update may deserve a deliberate pause.
- 02
Pause before the side effect
Place the boundary immediately before the consequential action. A review that starts after the external change can explain what happened, but it cannot prevent that change.
- 03
Design the evidence packet
Show the proposed action, the inputs that support it, the relevant recent browser state, and why review was requested. Prefer the smallest decision-ready packet over a wall of logs.
- 04
Name every valid decision
Define what approve, reject, and escalate mean for this workflow. Set a timeout or unavailable-reviewer path that defaults to no consequential side effect unless your team has explicitly chosen another safe behavior.
- 05
Keep the decision reusable
Attach the decision and short rationale to the source evidence. When the same edge case matters for later versions, treat the reviewed outcome as a candidate evaluation case rather than an isolated approval.
What success looks like
- The pause happens before one named consequential action.
- The reviewer receives enough context to decide without reconstructing the entire run.
- Approve, reject, escalate, timeout, and reviewer-unavailable behavior are explicit.
- Important reviewed outcomes have a defined path into later evaluation work.
If something is blocked
- If nearly every step needs approval, narrow the workflow or separate observation from intervention before adding more review gates.
- If reviewers repeatedly open raw logs, improve the evidence packet around the decision instead of asking them to investigate the whole session.
- If the timeout behavior is disputed, keep the side effect blocked until the workflow owner names a safe fallback.