Source: Don't Build Multi-Agents (Cognition, 2025)

Cognition (Devin) argues for single-threaded agents that share full context, warning that parallel subagents make conflicting implicit decisions.

Contents

Summary

Two principles for reliable long-running agents: share full context (whole traces, not just messages), and recognise that actions carry implicit decisions, so agents that can’t see each other’s decisions produce conflicting work.

Key claims

  • Example: splitting a Flappy Bird clone between a “background” subagent and a “bird” subagent yields mismatched pieces even when both get the original task.
  • Recommended: a single-threaded linear agent; for very long tasks, add a separate model that compresses history into key decisions and events (Cognition fine-tuned smaller models for this; hard to get right).
  • Claude Code’s subagents (as of mid-2025) mostly answer narrow questions rather than write code, and their value is keeping investigation out of the main context.
  • The 2024 pattern of a large model writing edit instructions and a small “apply” model executing them was fragile; one model now usually decides and edits in a single action.

Feeds

Coding agents

Suggest changes →