How to decide, in practice

Pulling Chapters 1 through 5 together into a decision guide:

• Is the task mostly reading (research, search, broad analysis)? Multi-agent with an orchestrator-worker pattern can pay off, if the value justifies the token cost and you invest in clean delegation.
• Is the task mostly writing (code, a single coherent document)? Prefer a single agent with unified context, and scale it with compression rather than parallelism.
• Is it mixed? Separate the phases: parallel reading, then single-threaded writing.
• Whichever you pick, the deciding factor for reliability is context engineering, supported by real infrastructure and outcome-based evaluation. If you go multi-agent, use MAST (Chapter 4) to find and fix your dominant failure modes.