Building Effective Agents

The useful distinction between a workflow and an agent is who chooses the next step. In Anthropic's agent-building guide, a workflow follows paths defined by software, while an agent chooses its tools and actions as the task unfolds. A predictable task can benefit from fixed stages with checks between them. An open-ended investigation may need the agent to change direction after seeing results. That flexibility adds cost and opportunities for errors to accumulate, so the guide recommends adding autonomy only when the task benefits from it and giving the system feedback and stopping conditions.

“Workflows are systems where LLMs and tools are orchestrated through predefined code paths.”

“Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”

“We suggest that developers start by using LLM APIs directly: many patterns can be implemented in a few lines of code.”

“While building our agent for SWE-bench, we actually spent more time optimizing our tools than the overall prompt.”