AI Agents Engineering · Orchestration & Multi-Agent · Lesson 2 of 3
Plan-and-execute, routing & handoffs
A few orchestration patterns recur across frameworks. Knowing them by name lets you pick the right shape instead of reinventing.
- Router: a classifier/model routes each request to the right tool, sub-agent, or workflow. Cheap, robust, great first move.
- Plan-and-execute: an orchestrator drafts a multi-step plan, then executes (often delegating steps to ReAct specialists), and can replan as results arrive.
- Orchestrator–workers: a lead agent decomposes and dispatches subtasks to workers, then synthesizes their results.
- Handoffs: one agent transfers control (and relevant context) to another specialist — the pattern behind the OpenAI Agents SDK.
- Evaluator–optimizer: a generator agent produces, a critic evaluates, and the loop iterates to a threshold.
Handoff hygieneThe failure point in routing/handoffs is context transfer. Pass the goal and the minimal necessary state — not the entire raw history — or the receiving agent gets confused and cost explodes.
◆ Lock it in
- Core patterns: router, plan-and-execute, orchestrator–workers, handoffs, evaluator–optimizer.
- Start with a router; add planning/handoffs only as complexity demands.
- Handoffs live or die on clean context transfer — pass minimal necessary state.
Feynman drill — say it out loudDescribe the plan-and-execute pattern and one situation where a simple router would be the better choice.
Step 1 rate your confidence · Step 2 pick your answer
In routing and handoff patterns, the most common failure point is:
Step 1 — how sure are you?