Reusable blueprints for agentic workflows: fixed steps + model-driven behaviour. Implemented with LangGraph and /chat/completions from first principles.
| Pattern | Essence |
|---|---|
| 1. Reflection | Generate → critique → refine |
| 2. ReAct Tool-Use | Reasoning + acting with tools |
| 3. Orchestrator–Worker | Central orchestrator + specialized workers |
| 4. Multi-Agent | Specialized agents with handoffs |
Generate → critique → refine. Answer generator produces a response; reflector gives feedback (tone, correctness, completeness). A decision node either loops back for another pass or ends. State: query, feedback, messages, answer, max_iterations.
Use cases: Code refinement (write → run → use errors as feedback), IR (retrieve → grade → refine).
Reasoning + acting. Chain-of-thought plus tool calls (search, calculator, APIs, optionally MCP). Model reasons, calls tools, gets results, repeats until it can answer or hits a cap. State: query, messages, reasoning, tool_calls, tool_call_responses, answer, iterations, session.
Use cases: External APIs via MCP, sandboxed DB queries.
Central orchestrator + workers. Orchestrator parses the query and emits sub-tasks; workers (e.g. Web-Search, Knowledge-Base) run in parallel via LangGraph Send; synthesizer aggregates results. State: query, workers, messages, result, session, final_answer.
Use cases: Multi-source QA (e.g. weather + calendar in one question).
Specialized agents + handoffs. Multiple agents, each with own prompt and tools; handoff is a tool that passes control to another agent. Graph cycles through the active agent; conditional edges route on handoff or end. State: iterations, session, messages, answer, query, tool_calls, handoff_message, current_state.
Use cases: Complex plans (e.g. 3-day Paris itinerary: calendar + travel + booking agents).
LangGraph provides State (TypedDict + reducers), Nodes, Edges, and Conditional edges (including fan-out via Send). Run with graph.invoke(input) or ainvoke. Patterns can be combined (e.g. Orchestrator–Worker with tool-using workers, or Multi-Agent with reflection per agent).
References: Phil Schmid – Agentic Pattern · LangChain multi-agent routing · Handoffs (customer support)