Multi-agent does not mean copying one error many times

Multi-agent work adds value only when specialization gains exceed the cost of communication, conflict and error propagation.

For critical facts, return to primary evidence and check from an independent context.

01 / TWO DIMENSIONSDefine sharing and isolation first

Shared context supports continuous reasoning but can contaminate every participant. Isolated context better fits specialist work but needs a clear communication protocol. A topology can be centrally managed, peer-to-peer or decentralized handoff. Before choosing it, ask the simpler question: would one agent with tools and staged workflow already be enough?

02 / WHEN IT HELPSUse multiple agents only for separable work

Multi-agent work suits tasks that can be split, run in parallel and mutually reviewed. One agent may extract facts, another check citations and a third check delivery format; each needs independent input and acceptance criteria before results are assembled. When every step depends on one long shared context, frequent communication can make the system slower than a single agent.

03 / COMMUNICATIONGive files, messages and permissions clear ownership

A shared filesystem invites concurrent overwrites, excessive messages consume context, and a manager agent can become the bottleneck. Define task claiming, file locks, result formats, deadlines and escalation rules. Every important conclusion needs its source and responsible agent, while authority to modify or send must match the task’s real boundary.

04 / FAILURE MODESStop cascades and false consensus

The most dangerous failure is an upstream error treated as fact by downstream agents and amplified. Mutual review can also create false agreement when everyone shares the same bias. Return critical facts to primary evidence for independent checking. Do not let multiple agents write one file at once; cross-check key steps in separate contexts; after repeated failure, return to human coordination instead of expanding the team. Final exercise: split case organization into facts, timeline, citations and delivery, then design message format, file ownership and final human acceptance.

Sources

bojieli/ai-agent-book.
Chapter 10 — Multi-agent collaboration.
This is an applied reorganization, not a replacement for the source book; use the source text and code for exact claims.

Return to the learning map and begin an experiment.

Open the learning map

Multi-agent does not mean copying one error many times.