Workflow

Workflow

Also called: 工作流

A predefined sequence of steps. Anthropic separates "workflows" (fixed code path, LLM works within nodes) from "agents" (the LLM directs its own process).

TaskWorkflow: fixed code pathAgent: LLM directs itself
Schematic (simplified)

Workflow vs agent

  • Workflow: the process is orchestrated by code; the LLM does local work at preset nodes (classify, extract, rewrite). Predictable, testable, cost-bounded.
  • Agent: the LLM decides what to do next, which tool to use, when to stop. Flexible, but harder to predict and debug.

Anthropic's advice: if a workflow solves it, don't reach for an agent; use agent autonomy only when the steps and branches cannot be enumerated in advance.

Common workflow patterns

Prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer. Each keeps a deterministic skeleton and calls the LLM only where needed. A workflow can also embed human approval points.

Common misconceptions

  • A workflow is not a "lesser" option — for most production cases it fits better than a fully autonomous agent.
  • Using an LLM inside a workflow does not make it an agent; the difference is *who decides the process*.

Related terms

Sources

  1. Anthropic — Building effective agents: workflows vs agents (2024)

Compiled 2026-08-29 · This glossary is compiled from public papers, official specifications, and common industry definitions, and is updated as the field evolves. Corrections welcome.