The contrarian essay from the Devin team that reset the industry's defaults.
Don't Build Multi-Agents
Walden Yan (Cognition) Jun 2025
Open cognition.com →Start with one. The famous industry debate (Cognition's "Don't Build Multi-Agents" vs Anthropic's multi-agent research system) actually agrees on the fundamentals: multiple agents working from partial context make conflicting decisions, and multi-agent setups burn roughly 15x the tokens. Go multi-agent only when the work is genuinely parallelizable and read-heavy, like broad research, and even then use one lead agent delegating to sub-agents, not a committee of equals.
18 resources.
The contrarian essay from the Devin team that reset the industry's defaults.
Walden Yan (Cognition) Jun 2025
Open cognition.com →The thread version of the argument, with the mistakes he keeps seeing.
Walden Yan Jun 2025
Open x.com →The honest update: some multi-agent setups now work, most sexy ideas still don't.
Walden Yan 2026
Open x.com →The follow-up naming the specific multi-agent patterns that survived production.
Cognition 2026
Open cognition.com →The other side of the debate: a 90% performance gain, at 15x the tokens.
Anthropic Jun 2025
Open anthropic.com →Anthropic's own decision rules for graduating beyond a single agent.
Anthropic 2025
Open claude.com →Shows why Cognition and Anthropic are both right about different problems.
Philipp Schmid (Google DeepMind) Jun 2025
Open philschmid.de →The architecture broken down with diagrams a non-researcher can follow.
ByteByteGo 2025
Open blog.bytebytego.com →200+ analyzed tasks and 14 failure modes, the evidence behind the caution.
UC Berkeley (MAST) Mar 2025
Open arxiv.org →The open dataset and annotator to diagnose your own multi-agent failures.
UC Berkeley Sky Computing Lab 2025
Open github.com →The digestible summary of where multi-agent systems break.
UC Berkeley Sky Computing Lab 2025
Open sky.cs.berkeley.edu →A neutral recap of the whole debate if you want it in ten minutes.
CTOL Digital 2025
Open ctol.digital →Frames the choice as a coordination-cost calculation, which is the right lens.
Cognilium 2026
Open cognilium.ai →Extracts the engineering lessons hiding inside Anthropic's post.
LLM Multi Agents 2025
Open llmmultiagents.com →Maps each MAST failure mode to a practical fix.
Future AGI 2026
Open futureagi.substack.com →The paper's findings retold for people who won't read a paper.
Anna Alexandra Grigoryan 2025
Open thegrigorian.medium.com →The case study written for someone deciding whether to copy the pattern.
ZenML 2025
Open zenml.io →The emerging alternative: one deep long-running agent instead of many shallow ones.
NVIDIA AI Podcast (Harrison Chase) 2026
Watch on YouTube youtube.com →The same ground, over in Build the product, our Starting Up track.