Hover a node · click to focus · ←/→ steps
Spec → allowlisted tools → model loop. Writes pause for human approval; MCP goes through an allowlist gateway.
When you need this
- Multi-step agent automation
- Human approval before external actions
- Scheduled agent maintenance jobs
- Need replay after worker crash
Prerequisites
- Temporal cluster or Temporal Cloud
- One bounded pilot workflow identified
- Approval channel (Slack / ticket) for humans
Tools
Strong for prod agents; assess ops learning curve vs simple queues.
Assess for complex agents; keep simpler RAG as plain pipelines until graph complexity pays off.
Steps
- 1
Pick a bounded pilot workflow
Choose a 3–5 step agent task with clear success criteria — not open-ended ‘do my job’ agents.
- 2
Model activities and approvals
Split into idempotent activities. Insert human approval signals before writes or external communications.
- 3
Compose agent logic carefully
Use LangGraph (or equivalent) for tool loops inside activities; keep Temporal as the durability and approval boundary.
- 4
Wire observability
Export workflow traces to OTel. Dashboard running, failed, and waiting-for-approval counts.
- 5
Document replay and failure modes
Test worker crash mid-workflow. Document how replay behaves and who unsticks approvals.
Adoption pitfalls
- Putting non-idempotent side effects in activities without dedupe keys
- No timeout on human approval → zombie workflows
- Monolithic agent that Temporal cannot meaningfully checkpoint
Adoption checklist
- Pilot workflow in production with owner
- Human approval on all external writes
- Replay tested after worker failure
- On-call runbook for stuck workflows
SEER REAL assessment / sprint
Assessment flags fire-and-forget agents with prod side effects. Sprint implements one durable workflow with approval gates and a crash-replay drill.