orchestrator-patternslisted

# orchestrator-patterns Owns the runnable **workflow shapes** for a single agent system. These five composable patterns (plus the ReAct loop) cover the large majority of real agent workflows; reach for a multi-agent topology only when none of them fit. ## When to use - Deciding how to structure steps: one prompt, a chain, a branch, a fan-out, or a refine loop. - You want a working reference implementation, not just a diagram. - Composing patterns (e.g. route → then chain inside one branch). ## Run it (every pattern is runnable offline) ``` python -m harness.orchestrator run --pattern sequential python -m harness.orchestrator run --pattern router python -m harness.orchestrator run --pattern parallel python -m harness.orchestrator run --pattern vote python -m harness.orchestrator run --pattern orchestrator-workers python -m harness.orchestrator run --pattern evaluator-optimizer ``` ## The patterns (detail in `reference.md`) 1. **Prompt chaining (`sequential`)** — decompose into fixed ordered steps; each step's output feeds the next. 2. **Routing (`router`)** — classify the input, dispatch to a specialized handler. 3. **Parallel — sectioning (`parallel`)** — split independent subtasks, run concurrently, join. 4. **Parallel — voting (`vote`)** — run the same task N times, aggregate (majority/best). 5. **Orchestrator-workers (`orchestrator-workers`)** — a planner decomposes dynamically, workers execute, a synthesizer combines. 6. **Evaluator-optimizer (`evaluator-optimizer`