An agent needs less freedom than you might think.

Anthropic distinguishes workflows, where code controls the path, from agents that choose their next steps. Its practical advice is to begin with simple, composable patterns and add autonomy when the task needs it. Routing, parallel work, and evaluator loops can solve many problems without an open-ended agent. More autonomy brings additional cost and opportunities for errors.

Anthropic · Dec 19, 2024

Why it matters

A useful lens for choosing which parts of an automated workflow should be predictable and which need model judgment.

More context

Start by asking whether the path can be described in advance. If it can, an explicit workflow is easier to inspect. For less predictable work, put clear tool interfaces and feedback around the agent. Evaluate whether that flexibility improves outcomes.

Keep in mind

This is practitioner guidance from a model provider, not proof that one architecture wins on every workload.

Original read