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.