AI Agents and Multi-Agent Systems · Lesson 4
Agent Architecture & Design Patterns
Common patterns: prompt chaining (fixed sequence), routing (classify then send to the right handler), parallelisation (run subtasks at once and merge), orchestrator-workers (a lead model splits work among workers), and evaluator-optimiser (one model drafts, another critiques).
Single-agent with many tools works for most tasks. Move to multi-agent only when tasks need separate expertise, parallel work or isolated contexts.
Start with the simplest pattern that works and add complexity only when evaluation shows it helps.
Key points
- Chaining, routing, parallelisation
- Orchestrator-workers for decomposable tasks
- Evaluator-optimiser for quality loops
- Prefer the simplest pattern that passes evals