Article

Why Do Multi-Agent LLM Systems Fail?

UC Berkeley (MAST) · arxiv.org

200+ analyzed tasks and 14 failure modes, the evidence behind the caution.

Open on arxiv.org ↗

The question it answers

  1. Should I build one agent or a team of agents working together?

Read with it

  1. The contrarian essay from the Devin team that reset the industry's defaults. cognition.com Article Don't Build Multi-Agents Walden Yan (Cognition)
  2. The thread version of the argument, with the mistakes he keeps seeing. x.com Thread Walden Yan: principles for building agents Walden Yan
  3. The honest update: some multi-agent setups now work, most sexy ideas still don't. x.com Thread Walden Yan: a year later, what changed Walden Yan
  4. The follow-up naming the specific multi-agent patterns that survived production. cognition.com Article Multi-Agents: What's Actually Working Cognition
  5. The other side of the debate: a 90% performance gain, at 15x the tokens. anthropic.com Article How we built our multi-agent research system Anthropic
  6. Anthropic's own decision rules for graduating beyond a single agent. claude.com Article When to use multi-agent systems (and when not to) Anthropic
eChai Partner Brands The partnership →