An agent architecture begins with deciding who controls the next step

Anthropic’s guide separates workflows, where code determines the sequence, from agents, where a model chooses its next actions and tool calls. That distinction is more useful than applying the same label to every multi-step AI system.

The article lays out patterns including prompt chains, routing, parallel work, dynamically assigned subtasks and iterative evaluation. Each fits a different structure of uncertainty. A fixed sequence can be effective when the task is predictable; model-directed work becomes useful when the needed steps cannot be specified in advance.

Autonomy also adds cost, latency and opportunities for errors to compound. Agents need feedback from the actual environment and clear stopping conditions. Tool definitions deserve the same design attention as prompts: Anthropic describes eliminating file-path mistakes by requiring absolute paths.

The page now notes that its tooling landscape has changed since publication. Its durable material is the architectural reasoning and emphasis on measurable improvement, rather than treating more elaborate orchestration as an achievement by itself.