Learning path · Agents & Orchestration · 63
Multi-Agent Systems
Multiple specialized agents coordinating—researcher, coder, reviewer—via shared state or message passing.
Why it matters
- Mirrors team workflows but adds coordination overhead.
- Framework choice affects debuggability and cost.
- Needs clear role boundaries to avoid duplicate work.
Key ideas
- Role specialization
- Orchestration
- Shared memory
Top resources
- 01ArticleAnthropic
How we built our multi-agent research system
Why this resource. A real multi-agent research system and its coordination costs.
Covers in this concept
- subagents
- orchestration
- 02DocsLangChain
LangGraph
Why this resource. How to encode multiple actors as a graph rather than a chat pile.
Covers in this concept
- state graph
Multi-agent setups divide labour: one agent gathers facts, another drafts, a third verifies policy. Wins on complex deliverables; fails when roles blur and agents talk past each other. Define interfaces—typed messages, shared blackboard, supervisor graph—and measure end-to-end success, not per-agent eloquence. Prefer fewer agents until single-agent plus tools plateau. Start with one agent plus tools; add specialized agents only when traces show repeated context overload on a single thread.
Updated 2026-08-09 · Full learning path