What are agent workflow patterns?
Agent workflow patterns are the small set of reusable shapes engineers use to compose LLM calls, tools, and control flow. Anthropic's canonical list covers prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer. Most production agents stitch two or three together rather than running one open-ended loop.
How the patterns fit together
Agent workflow patterns are the shapes engineers compose when wiring LLMs, tools, and control flow. Most production systems mix two or three simple patterns instead of reaching for a monolithic agent. Anthropic's Building Effective Agents post is the canonical taxonomy.
The patterns
Prompt chaining. Serial LLM calls where each consumes the prior output, with optional gates between steps. Good for fixed pipelines like draft, then critique, then translate.
Routing. A classifier LLM directs the request to one of N specialized branches. Pairs well with a cheap model out front and expensive models behind the right gate.
Parallelization. Fan-out either by voting (same task, N runs, aggregate) or sectioning (split into independent subtasks). Voting raises confidence on safety checks; sectioning shortens wall time.
Orchestrator-workers. A planner LLM spawns subagents at runtime for tasks it cannot pre-list. Better than chaining when the shape of the work is unknown until inspection, like multi-file code edits.
Evaluator-optimizer. A second LLM critiques the first's output and loops until acceptance criteria pass. Used for translation, deep search, structured extraction.
What to watch for
Underneath sits ReAct: reason, act, observe, repeat. Frameworks like LangGraph, CrewAI, and Inngest expose these as graph nodes or steps you compose. The non-obvious bit: multi-agent debate often underperforms a single well-prompted call with an evaluator pass. Reach for orchestrator-workers only when subtasks are genuinely unknowable upfront.
For workflow migrators
Chaining, routing, and sectioning map cleanly to Zapier or n8n branching. Orchestrator-workers and evaluator-optimizer need an agent runtime since the next step isn't known at design time.
Last updated: May 20, 2026