# Workflows vs. Agents

**Track:** Models, Tools & Agents — AI-Native Full-Stack Development — complete (30)
**Framework / surface:** web dev (orchestration architecture)
**Level:** Intermediate
**Prerequisites:** The Agent Loop, Tool Calling: Giving Models Capabilities

**In one line:** Use predefined steps when the path is known; spend agentic autonomy only where choosing the next step creates value.

## Theory, aesthetics & inspiration

Not every sequence containing a model call should be an agent. Anthropic's "Building Effective Agents" (December 2024) drew the line that stuck: workflows orchestrate models and tools through predefined code paths, while agents direct their own process, choosing which tools to use and how many steps to take. Autonomy is a cost — more model calls, more variance, harder testing, new security surface — so the design rule is conservative: encode known structure in ordinary code, and spend model judgment only on decisions that cannot be specified economically in advance. A workflow can still contain models, branching, retries, and parallelism; an agent is earned when choosing the next step is itself the value. The boundary also moves: a task that needs an agent today becomes a deterministic workflow tomorrow, once its successful trajectories are understood — and building both around explicit tools, schemas, and evals is what keeps that promotion cheap. The goal is never maximum autonomy; it is the simplest control structure that reliably does the job.

**Builder question:** Does this task require judgment about the next step, or only execution of steps the application already knows?