Skip to main content

Links

2026

Don't Build Multi-Agents — Cognition

·39 words·1 min
The Devin team’s case against splitting work across agents: context lost between agents costs more than the parallelism buys. Read it together with their follow-up, Multi-Agents: What’s Actually Working, where they concede multi-agent setups work when writes stay single-threaded.

Building Effective AI Agents — Anthropic

·29 words·1 min
Still the best starting point on agent architecture. The workflows-versus-agents distinction and the five patterns — prompt chaining, routing, parallelization, orchestrator-worker, evaluator-optimizer — became the vocabulary everyone builds with.