Agent safety testbed
Work a simulated incident with a browser agent. The human can restrict services and capabilities while the browser tool surface updates live.
What to try
Pick a scenario, assign the agent an investigation boundary, challenge its hypothesis, then approve or reject its request.
WebMCP is not detected in this browser.
Use ChatGPT Desktop's in-app browser, or Chrome 149+ with the local testing flag enabled.
chrome://flags/#enable-webmcp-testing
The simulator still works here, but the browser agent will not receive tools. From DevTools you can inspect the fallback map with:
await incidentCommandTools.get_incident_state.execute({})
await incidentCommandTools.investigate_incident.execute({ serviceId: "checkout" })
Error rate
p95 latency
Affected sessions
Revenue risk
Human Console
Inspect evidence, mark a suspect, and steer the agent.
Hypothesis
Facts and agent reasoning stay separated.
Proposed Action
Risky actions require explicit approval.
Decision Context
Review the evidence and exact production scope before deciding.
Human Approval Required
Production changes are gated here; rejection needs a reason.
Run Scorecard
The incident is closed and the agent run has been graded.