Key takeaways

• Single agent: sequential, unified context, simple and reliable, but slow and limited by the context window.
• Multi-agent: parallel, specialized, broad, but costly (around 15 times the tokens of a chat), harder to coordinate, and harder to debug.
• The decisive question is read versus write, not single versus multi. Read tasks parallelize well; write tasks need unified context.
• Mixed tasks should split the read phase from the write phase architecturally.
• Whatever you choose, context engineering, good infrastructure, and real evaluation decide whether it works.

Continue to Chapter 2 for the strongest real example of a multi-agent system done well.