Volume III
Single-Agent Patterns
Workflows that earn their extra hop
Most problems do not need an autonomous loop. Anthropic's catalog — chaining, routing, parallelization, evaluator-optimizer — is the correct starting vocabulary. Add a free-running agent only when the path cannot be named in advance.
- III.1Prompt ChainingA sequence of LLM calls where each step's output is the next step's input, with programmatic gates between them. The first workflow to reach for.5 min
- III.2RoutingClassify the input, then send it to a specialist prompt, model, or workflow. Separation of concerns for language.5 min
- III.3ParallelizationFan-out independent LLM work, then reduce. Sectioning for speed; voting for diversity. Aggregation is code, not another creative writer.5 min
- III.4Evaluator–OptimizerOne model generates, another critiques against explicit criteria, loop until the critic accepts or the budget dies. The missing unit test of generation.6 min
- III.5ReAct and Tool UseThought, action, observation. The original agent pattern. Still the right inner loop — as long as 'thought' is optional and 'action' is schema-bound.6 min
- III.6Code-ActThe model writes a program to use tools, instead of emitting one JSON call at a time. Hugging Face smolagents made this the default for a reason: code composes.6 min