Key takeaways

• The single-versus-multi "debate" is a false dichotomy; both Anthropic and Cognition are right for their tasks.
• Both positions reduce to the same core: context engineering decides reliability.
• Cognition avoids dividing context; Anthropic divides it but engineers the division to be clean. Same problem, two valid strategies.
• Task shape decides architecture: read tasks favor multi-agent parallelism, write tasks favor a single unified-context agent.
• The hardest part for both is the last mile from prototype to production, and the common foundation is the quality of the agent's context.

This completes the Multi-Agent chapters. For the context engineering ideas these chapters lean on, see the context engineering folder; for memory, see the agentic memory folder.