PROJECT FIVE - MULTI-AGENT AND PRODUCTION HARDENING

LEARNING GOALS By the end of this chapter you will be able to split one agent into a supervisor and two subagents. You will be able to route cheap work to a cheap model tier and careful work to a stronger one. You will be able to add retries with exponential backoff and a maximum-iteration safeguard. You will be able to fill in an honest production-readiness checklist for a finished system. FROM ONE COOK TO A SMALL KITCHEN Until now one agent did everything: it read, judged, and wrote. A small restaurant kitchen works differently. A prep cook chops quickly and cheaply, a head chef finishes the plate with care, and an expediter at the pass decides who gets which order and checks that nothing goes out wrong. In multi-agent terms, the expediter is the supervisor, which is an agent that hands tasks to other agents and collects their results. The cooks are subagents, which are agents that each do one narrow job. Splitting work this way brings three benefits. Context isolation means each subagent sees only what its job requires, so its desk stays clean. Specialization means each subagent can have its own instructions, tools, and model. And cost control means you pay for a powerful model only where power is needed. It also brings costs of its own, namely more moving parts and more places for a handoff to go wrong, so use this pattern only when it earns its keep.