CH 03 · The Case Against: Cognition's "Don't Build Multi-Agents" · 05/10
Read as one pagePrinciple 1: Share context, and share full traces
The obvious patch is to give the subagents the original task as context. Yan agrees this helps, and states it as the first principle:
Principle 1: Share context, and share full agent traces, not just individual messages.
The emphasis on full traces matters. In a real production system the conversation is multi-turn, the agent probably made tool calls to decide how to break down the task, and any of those details could change how a subtask should be interpreted. Sharing just the final subtask instruction is not enough; a subagent needs to see the whole trajectory that led to its assignment, or it will fill the gaps with its own assumptions.
Finished this idea?
Your learning progress stays in this browser.