AI Adoption · Engineering Playbook
Pattern · Architecture

Multi-agent debate

Anthropic vs Cognition

W
← All patterns Source: research/patterns.md

Anthropic: "How We Built Our Multi-Agent Research System" — orchestrator-worker pattern, +90.2% over single-agent on internal evals, but ~15× cost.

Cognition: "Don't Build Multi-Agents" → "What's Actually Working" — the only safe pattern is reads fan out, writes stay single-threaded.

FRAGILE · multiple writers A1 A2 A3 shared write target conflicts · poor context sharing SAFE · reads fan out, write single-threaded R1 R2 R3 trunk single writer
The resolution of the multi-agent debate. Many agents read, one writes. Almost every failed multi-agent prototype violates this.
Default to single-threaded with subagent reads. Multi-agent writes only when work is parallel and conflict-free. Always run a cost ratio before recommending.