liveBook · Manning

Multi-agent system

As explained in Designing AI Agents

A software system in which multiple agents coordinate or work independently to perform tasks.

A multi-agent system uses multiple agents to perform parts of a task, either independently or through coordination. Adding agents can provide fresh context, specialized work, parallel search, or additional viewpoints, but each participant creates another context window, reasoning chain, and communication channel. section 2.4.3 section 8.0

Start with one agent when it can meet the task’s acceptance criteria. Add another only when it contributes work, evidence, tool access, authority, or judgment that the first context lacks—and when its contribution can be reconnected to the original goal. Independent, decomposable work is a stronger candidate for delegation than tightly sequential work. section 8.0 section 2.4.3

Parallel company research

A lead agent divided a broad company search among subagents, each responsible for a bounded set of companies. The lead gathered their findings, tracked which sets had returned, checked citations, and identified missing coverage. Parallel work helped because the company sets could be searched independently. section 8.0

A common middle ground is a primary agent that owns decomposition, shared state, and final integration while bounded subagents execute assigned work. This preserves integration authority while providing fresh contexts or parallel execution. section 2.4.3

Account for the costs at agent boundaries:

  • Coordination: handoff summaries can omit context, sequential handoffs add latency, and additional context windows require loading and monitoring. section 2.4.3
  • Error amplification: parallel agents may pass unchecked errors to an aggregate. One study of 260 configurations found amplification of up to 17.2 times a single-agent baseline for independent multi-agent systems. section 2.4.3
  • Debugging: failures require reconstructing message flow, state ownership, and the order in which partial results affected later decisions. section 2.4.3

Architecture benefits depend on task shape. In one study, every tested multi-agent architecture worsened performance on a sequential-planning benchmark, while centralized coordination improved performance by 80.8% on a decomposable financial-reasoning task. Compare the proposed design against a single agent using task success, latency, token cost, and debugging effort. section 8.0 section 2.4.3

How Designing AI Agents covers it

How deep it goes
Explains coordination costs, failure modes, and primary-agent ownership while deferring detailed handoff, fan-out, and hierarchy patterns.
Examples
Illustrates parallel company research and compares results on sequential-planning and decomposable financial-reasoning benchmarks.
What you can do after
Evaluate whether to add agents using task success, latency, token cost, and debugging effort.
How it connects
Connects multi-agent systems to task decomposition, parallel execution, shared state, context windows, and orchestration.

Also explained in

Compare how 13 books explain multi-agent system →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Designing AI Agents. The section links point to the parts of the book it draws on.