Skip to content

Recipes: closed real-world cases

A recipe is a real business case from start to finish — not a feature demo. Each one follows the same template so you can compare cases and lift the structure for your own:

  1. The business problem — what the system is for, who uses it.
  2. Graph design — the topology (and why it is shaped this way).
  3. State schema — what flows between nodes.
  4. Durability & HITL — why this case needs checkpoints / interrupts.
  5. Testing strategy — how it is verified offline, no LLM.
  6. Cost & observability — what to watch, what it costs.
  7. Deploy — how it runs in production.
  8. How to adapt — the knobs you turn for a different business.

Each recipe points at a runnable example in examples/ so you are never reading abstract code.

Recipe Business case Runnable example Core pattern
FastAPI agent in 10 minutes Durable AI agent over HTTP fastapi_server + teff/scaffold/fastapi Flow.react() + checkpointer + SSE streaming — the zero-to-30s on-ramp
Fraud review Payment screening fraud_gate Dynamic Command routing + HITL + durable resume
Release approval Ship-gate with human sign-off release_coordinator Supervisor + route() + Map + approval gate
Support triage Customer-support SLA hub simple_router + repair-ai-chat route() supervisor + RAG + HITL escalation
Ops daemon Scheduled change-triage repo-health YAML-only agent-driven tick via CLI daemon
KB assistant Grounded self-service Q&A rag_search + repair-ai-chat RAG retrieval + confidence gate + human fallback
Invoice processing AP automation with approval structured_output + agent_approval Extract/Fallback + Validate + durable approval

Reading the pattern

The six cases differ in business, but the graph anatomy repeats:

  • One entry point — a node that builds context (ingest / reset / context builder).
  • A driver — either an agent loop (react_agent, ReActAgent) or a supervisor (route() / Supervisor) that decides what happens next.
  • A fan-out — Parallel (fixed branches) or Map (runtime list).
  • A human gate — Interrupt / Ask where a wrong answer is costly.
  • A finalizer — a node that commits, sends, or writes the outcome.

If your case is a mix, combine the recipes — e.g. a support hub that also re-triages nightly (triage + ops daemon) or a release gate with a fraud check (approval + fraud).