OpenAI Build Week · Developer Tools

Incident triage with evidence, boundaries, closure

A sanitized SRE copilot demo that transforms synthetic alerts, logs, traces, metrics, and change history into an auditable evidence package. The assistant can reason and recommend, while production actions stay behind explicit human approval.

Run RCA mission All signals are built-in mock data

RCA mission control

Agents investigate a common production topology

A synthetic incident moves through serial and parallel investigation phases. Specialist agents collect read-only evidence from a typical company stack, then the control plane gates any production-changing action behind human approval.

Scenario selector

Switch the incident, keep the same control plane

Watch specialist agents sweep the topology for read-only evidence.

Active phase

Initializing RCA mission

Loading synthetic company topology and agent workflow.

Simulation control

One-click RCA demo run

Idle · scripted simulation · synthetic data · no live LLM or production tools.

Confidence model

Evidence coverage

-- evidence coverage · capped below 100%

Measures how complete the evidence bundle is. Each candidate cause carries its own hypothesis likelihood.

Approval boundary

Allowed checks vs blocked actions

Runbook retrieval mock

Read-only advisory snippets

Synthetic runbook context is retrieved for realism. It is advisory context, not proof.

Approval gate interaction

Recommended, not executed

Blocked actions can generate a local packet. No request is submitted anywhere.

Generated RCA brief

Copyable incident handoff

Approval packet

Local-only mitigation request

Platform blueprint

From noisy signals to controlled action

Select a layer to inspect its role in the synthetic incident workflow. Model reasoning is separated from deterministic policy, evidence tracking, and approval.

Guided RCA replay