OpinionProvider guide
Don't build multi-agents: share full context and keep decisions in one thread
An essay from Cognition, which makes Devin, argues that agents splitting a task among parallel subagents were fragile in 2025, and sets out two principles of context engineering instead.
Documented
What the piece reports as fact.
- As of June 2025, Claude Code started subtask agents but never ran them in parallel with the main agent, and mostly asked them to answer a question rather than write code.
- In 2024 many coding tools, Devin included, used an edit apply model: a large model described an edit in markdown and a small model rewrote the whole file. The essay says one model now usually decides and applies the edit.
- Cognition has fine-tuned a smaller model to compress an agent's history of actions and conversation into its key details, events and decisions.
- OpenAI's Swarm and Microsoft's AutoGen libraries promote multi-agent architectures.
Interpretation
Cognition's argument, in its own terms.
- Context engineering is the most important job of engineers building AI agents.
- Principle 1: share context, and share full agent traces, not just individual messages.
- Principle 2: actions carry implicit decisions, and conflicting decisions carry bad results. Rule out by default any architecture that breaks either principle.
- A single-threaded linear agent gets you very far. In 2025, running several agents in collaboration only results in fragile systems.
Dossiers
- Devinthis guide's dossier