Anthropic's Building effective agents draws a sharp line:
- Workflows — LLMs and tools are orchestrated through predefined code paths.
- Agents — LLMs dynamically direct their own processes and tool use.
Prefer workflows when
- The task decomposes into known steps (extract → classify → route).
- Latency and cost predictability matter.
- Failures need to be traced to a specific node.
Prefer agents when
- The path is unknown at design time (open-ended research, triage across novel data).
- The model can reasonably recover from its own mistakes.
- You can afford variable latency and token spend.
Common patterns
Prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer loops are all documented with code in Anthropic's agents cookbook. Start there before reaching for a framework.