The insight that reframes the whole debate: read versus write

Here is the idea that the comparison article and the architecture-debate article both land on, and it is the most useful takeaway in this entire folder. The important distinction is not single versus multi. It is whether your task is mostly reading or mostly writing.

• Read tasks are about gathering and analyzing information: research, search, fact-finding, synthesis. These split naturally into independent parts, so they parallelize well and suit multi-agent systems.
• Write tasks are about producing one coherent artifact: code, a long document, a set of file edits. These parts depend on each other, so splitting them across parallel agents creates coordination problems. They favor a single agent with unified context.
• Mixed tasks should be separated by phase: do the reading (research) in a parallel, multi-agent way, then do the writing in a single, unified-context way.

This is why, as the next chapters show, Anthropic chose multi-agent for a research tool (a read task) and Cognition chose single-agent for a coding tool (a write task). They did not disagree about principles. They were solving differently shaped problems.