Multi-agent systems

D1 · StoryFor young children and total beginners

Think of a team project. Who plans, who does the work, who checks?

Big jobs are easier with a team: one person plans, others do the parts, and someone checks the work.

Multi-agent systems are teams of AIs.

Like: A team with a planner, workers and a checker.

D2 · PictureFor curious beginners

An orchestrator agent splits a task and hands pieces to sub-agents, each with its own instructions and context window.

Sub-agents return summaries rather than everything they saw, which keeps the orchestrator's context focused.

Benefits: parallel work and separation of concerns. Costs: more tokens, coordination errors and harder debugging.

Like: A team with a planner, workers and a checker.
Where it breaks: Human teammates share background knowledge. Sub-agents share nothing except what's written in their instructions, so vague hand-offs fail silently.

D5 · FrontierFor experts

Splitting work across agents adds parallel effort, but also hand-offs, cost and new ways to fail.

Cost. Anthropic measured its own multi-agent research system using about 15 times the tokens of a chat (a single agent about 4 times), worth it only for valuable tasks that split well (Anthropic, 2025).

Why they fail. An analysis of over 1,600 traces from seven frameworks found 14 failure modes, in system design, misalignment between agents, and checking results (MAST, 2025).

Errors spread. In simulation, one adversarial image in one agent's memory spread a jailbreak to nearly all of up to a million agents (Agent Smith, 2024).

Trust between agents. Protocols such as A2A let agents from different vendors work together without exposing their internal tools or memory (A2A specification), so trust rests on identity and permissions rather than inspection.

Examples as of 2026-10-30

• Claude Research (a lead agent with subagents)
• Microsoft Agent Framework (successor to AutoGen)
• CrewAI
• LangGraph multi-agent workflows

Sources

1. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation (2023)
2. How we built our multi-agent research system (Anthropic) (2025-06-13)
3. Why Do Multi-Agent LLM Systems Fail? (2025-03-17)
4. Agent Smith: A Single Image Can Jailbreak One Million Multimodal LLM Agents Exponentially Fast (2024-02-13)
5. Agent2Agent (A2A) Protocol Specification (ongoing)

How it connects

• The agent loop runs inside each Multi-agent systems.

More in Layer 8 · Agents

• The agent loop
  Plan, act with a tool, observe the result, and repeat until the goal is met.
• Agent harnesses
  The program wrapped around a model that turns it into an agent: it runs tools, asks permission, manages context and keeps memory.
• Coding agents
  Agents that work inside a code repository: read files, make edits, run commands and tests, and iterate until the checks pass.
• Computer & browser use
  Agents that operate a screen like a person: look at a screenshot, then click, type and scroll.