Anthropic says most teams over-engineer their AI agents. Here's what actually works.

Anthropic's engineering team published a guide that goes against the hype: don't build complex agent architectures unless you have to. Start with the simplest solution and only add complexity when it demonstrably improves outcomes.

They draw a crucial distinction between workflows (LLMs orchestrated through predefined code paths) and agents (LLMs dynamically directing their own processes). Workflows give you predictability. Agents give you flexibility. Most teams reach for agents when a workflow would do.

Five production patterns cover most use cases: prompt chaining (sequential steps), routing (classify input, send to specialist), parallelization (run subtasks simultaneously), orchestrator-workers (central LLM delegates dynamically), and evaluator-optimizer (one LLM generates, another critiques).

The counterintuitive advice: skip the frameworks at first. Most of these patterns are just a few lines of API code. Frameworks can obscure the underlying prompts and make debugging harder.