2.3.1. Connected Agents, Handoffs, and Coordinator Patterns
💡 First Principle: Orchestration patterns differ in who decides what happens next. In a coordinator (orchestrator) pattern, one lead agent routes subtasks to specialists and assembles results. In a handoff pattern, control transfers from one agent to another that owns the next stage. Choosing the pattern is choosing where the routing logic lives.
A coordinator pattern suits tasks that fan out and recombine (research several aspects, then synthesize). A handoff pattern suits pipelines where each stage fully owns its phase (intake → diagnosis → resolution), passing the baton forward. Connected agents let one agent invoke another as if it were a tool, which is the building block for both patterns.
⚠️ Exam Trap: A coordinator that holds all the tools itself and just "thinks harder" is not multi-agent — it's one overloaded agent. Genuine orchestration distributes tools to specialists. If an answer keeps every capability on one agent, it's missing the point of the pattern.
Reflection Question: Why does giving each specialist a narrow toolset improve tool-selection accuracy compared to one agent holding every tool — and how does that connect to the "more tools" misconception?
