Module 4 of 6 · Advanced · 8 min read
Architectures and orchestration
Chaining, routing, parallelisation, orchestrator-workers and evaluator loops.
- 1Start with a single agent; add structure only when it fails.
- 2Orchestrator-worker setups suit broad tasks like research.
- 3Multi-agent systems use many more tokens: make sure the task is worth it.
01The pattern ladder
Climb only as high as you need. Each step adds capability, cost and complexity.
- Single call with good context.
- Prompt chain: fixed sequence of calls with checks between.
- Routing: classify, then send to a specialised handler.
- Parallelisation: split independent subtasks, or vote.
- Orchestrator-workers: a lead agent plans and delegates.
- Evaluator-optimiser: generate, critique, revise.
- Autonomous agent: open-ended loop with tools.
02When multi-agent helps
Breadth-first tasks, such as researching many sources or reviewing many files, benefit from parallel sub-agents, each with a clean context. Anthropic's research system and many coding agents use this pattern.
Tightly coupled tasks, such as editing one complex function, usually work better with a single agent, because coordination overhead outweighs the benefit.
03Briefing sub-agents
Sub-agents fail when briefs are vague. Give each one an objective, the output format, the tools to use and clear boundaries, exactly as you would brief a contractor.
Finished reading?
Mark it complete to track your progress, then test yourself with the AI Agents quiz.