Skip to content

Design Patterns

Design patterns are complete, runnable workflow definitions for common orchestration problems. Each page takes one problem, such as parallel fan-out, sagas, timers, or human approval, and gives you a working definition to register, run, and adapt to your own tasks. This section covers workflow patterns. Agentic patterns and agent recipes live in AI Cookbook.

Services Events AI & LLMs Cookbook recipe JSON or code Durable run
Microservice orchestrationHTTP service chains, conditional branching, parallel HTTP calls with Fork/Join. Dynamic parallelismDynamic forks, per-branch tasks, same-task fan-out, and parallel sub-workflows. Wait and timer patternsFixed delays, scheduled execution, external signals, and human approvals. Sending signals to workflowsComplete a blocked wait from an approval UI, webhook, or external callback. Task timeouts and retriesExponential backoff, leases, hard SLAs, and thundering-herd prevention. Saga and compensationUndo a partially completed transaction: failureWorkflow, reverse-order rollback, idempotent undo. Polling a long-running jobWait on a slow third-party API with one HTTP_POLL task, backoff, and a poll ceiling. Scheduled workflowsCron execution, downtime catchup, bounded windows, and concurrent runs. Event-driven recipesKafka, NATS, RabbitMQ, and SQS events that start or advance workflows. AI CookbookAgentic workflows and SDK-authored agents: RAG, MCP tools, guardrails, human approval, deep research, handoffs, and memory. Dynamic workflows as codePython-defined chains, branches, parallel work, loops, and generated definitions.