Skip to slide
25 / 41

CH 03 · The Case Against: Cognition's "Don't Build Multi-Agents" · 10/10

Read as one page

Key takeaways

  • Cognition's argument rests on reliability: long-running agents must contain compounding errors, and context engineering is the core of doing so.
  • The naive parallel-subagent pattern is fragile, because subagents misread subtasks (the Flappy Bird example) and make conflicting hidden decisions.
  • Principle 1: share full agent traces, not just individual messages.
  • Principle 2: actions carry implicit decisions, and conflicting decisions produce bad results.
  • The recommended default is a single-threaded linear agent; for very long tasks, add a hard-to-build compression step rather than splitting into parallel agents.
  • Claude Code uses subagents only for context isolation on well-defined questions, never for parallel code-writing, exactly because of the conflicting-decisions risk.

Continue to Chapter 4 to see the research on why multi-agent systems fail in practice.

Finished this idea?

Your learning progress stays in this browser.