In this post, we discuss the key learnings from our work with various teams on large language model (LLM) agents, emphasizing the success of simple and composable patterns over complex frameworks. We provide practical advice for developers on building effective agents, distinguishing workflows from agents and exploring commonly used agentic systems such as prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer. Additionally, we delve into the emerging concept of agents, which are autonomous LLMs that engage in planning, reasoning, and tool usage independently. We also provide insights on when to use agents, with examples from customer support and coding agents. Finally, we highlight best practices for tool development to enhance agent performance and reliability.
https://www.anthropic.com/research/building-effective-agents