AI Agents — Book 3 · The Five Patterns

Almost every useful agentic system is one of five shapes, or two joined together, and knowing the names is worth more than it sounds — it turns "I need an AI thing that handles enquiries" into "that is routing with a chain behind each branch", which is something you can build, cost and test. Prompt chaining, routing, parallelisation by sectioning or voting, orchestrator-workers, and evaluator-optimiser. Which four of the five are workflows rather than agents, and why that means they can be tested instead of merely evaluated. Ends with a four-question decision procedure you run before writing anything.