Orchestration patterns
Runbooks, dashboards, reviews, and handoff material make the work auditable.
What changes
Supervisor-worker
A planner or coordinator decomposes work and dispatches subtasks to specialized workers. Results are collected and synthesized before a human or downstream system acts.
When to use: research plus writing plus review, triage plus enrichment plus routing, code inspection plus test planning, or document processing with validation.
What we define:
- subtask queue and dependency ordering
- worker assignment criteria
- structured handoff schema
- result aggregation rules
- partial-failure handling
- review and approval point
What changes
Parallel pipelines
Independent branches run concurrently and merge at a synchronization point.
When to use: processing multiple sources, dataset chunks, customer records, repositories, or independent analyses.
What we define:
- concurrency limits
- per-branch timeout and retry policy
- merge strategy
- failure threshold
- cost budget per run
- audit output
What changes
Sequential chains
Steps execute in order and pass structured output forward.
When to use: extraction → validation → enrichment → classification → storage, or incident intake → context gathering → summary → suggested actions.
What we define:
- typed step inputs and outputs
- idempotency rules
- validation checks
- retry boundaries
- manual correction path
What changes
Human-in-the-loop
Workflows pause at defined checkpoints for approval, correction, or escalation.
Common checkpoints:
- sending an external message
- writing to a production system
- opening or merging a pull request
- changing access, infrastructure, or billing state
- accepting low-confidence extraction output