# Workflow vs Agent

Workflows follow predefined code paths; agents let the model dynamically direct its own steps and tool use.

## Why the CCAR-P exam tests this

CCAR-P tests choosing the least-autonomous pattern that solves the problem, since unneeded agency adds cost, latency, and failure modes.

## Definition

A **workflow** orchestrates LLM calls and tools along a **predefined path** written in code — prompt chaining, routing, parallelization. The control flow is fixed; the model fills in content at known points. This is predictable, testable, and cheap.

An **agent** hands control flow to the model: it decides which tools to call and in what order inside an agentic loop, adapting to results you couldn't script ahead of time. This buys flexibility for open-ended tasks but costs determinism.

The design rule is **start simple**. Use a workflow when the steps are known and stable; reach for an agent only when the task genuinely requires runtime decisions the code can't enumerate. Many 'agent' requirements are actually a routing or chaining workflow in disguise. When in doubt, prefer the more constrained pattern and add autonomy only where it earns its keep.

## Exam trap

Picking the agent 'because it's more capable' — the exam rewards the simplest sufficient pattern, and defaulting to full autonomy where a fixed workflow would do is the wrong answer.

## Commonly confused with

- **agentic-loop** — The agentic loop is the mechanism inside an agent; workflow-vs-agent is the higher-level choice of whether to use that mechanism at all.
- **single-vs-multi-agent** — This chooses workflow vs any agent; single-vs-multi picks how many agents once you've committed to agency.

## Related

- agentic-loop
- orchestrator-worker-topology
- single-vs-multi-agent
- batch-vs-synchronous
- model-selection

---

Source: Claude Cert Prep — an independent, unofficial CCAR-P study resource (domain P1). Not affiliated with Anthropic.