Type: url
Original: https://cognition.ai/blog/dont-build-multi-agents
Status: partial
Summary
A June 2025 Cognition blog post arguing that multi-agent architectures, in which subagents work on parts of a task in parallel, are fragile for long-running production agents. It sets out two principles of context engineering — share context, including full agent traces, and recognise that actions carry implicit decisions — and recommends a single-threaded linear agent, optionally with a model that compresses history, using Claude Code’s subagents and edit-apply models as examples.