CH 03 · The Case Against: Cognition's "Don't Build Multi-Agents" · 01/10
Read as one pageThe Case Against: Cognition's "Don't Build Multi-Agents"
Cognition argues long-running agents need one unified context to contain compounding errors, since split work across subagents causes conflicting decisions.
10 slides · 6 min read
In this chapter
- 01Why think in principles at all
- 02The foundation: context engineering and reliability
- 03The fragile architecture, illustrated with Flappy Bird
- 04Principle 1: Share context, and share full traces
- 05Principle 2: Actions carry implicit decisions
- 06What to build instead
- 07Real-world examples Yan offers
- 08On multi-agent collaboration specifically
- 09Key takeaways
This chapter explains Walden Yan's article "Don't Build Multi-Agents," written from Cognition, the team behind the Devin software engineering agent. Where Chapter 2 made the case for multi-agent systems, this article argues that for serious, long-running, reliable agents, you should usually avoid them. The two are not as contradictory as they sound, which Chapter 5 resolves, but first it is worth taking this argument on its own terms, because it is sharp and well reasoned.
Finished this idea?
Your learning progress stays in this browser.