Recapping each side fairly

The article gives a fair summary of both, which is worth restating because it makes the reconciliation click.

Anthropic's research agent uses the orchestrator-worker pattern. A lead agent (Claude Opus 4) plans and spawns subagents (Claude Sonnet 4) that search different facets of the query in parallel, then it synthesizes their findings into a cited report. It works because the search-and-retrieve step parallelizes naturally, and because, in Anthropic's own words, multi-agent systems "work mainly because they help spend enough tokens to solve the problem." Parallelizing breaks past the context window and sequential-speed limits of a single agent, letting the system comb through hundreds of sources in minutes. The cost is real: roughly 15 times the tokens of a chat, plus three kinds of complexity (coordination, debugging, and engineering).

Cognition's Devin uses a single agent. Its case is that multi-agent systems are fragile because partitioning context causes miscommunication: subagents miss the nuance needed to complete subtasks correctly. From this come the two principles from Chapter 3 (share full context, and beware conflicting implicit decisions). Rather than orchestrating multiple agents, Cognition scales a single agent's reach through context compression. McGuinness finds this convincing on its own terms: simpler designs are more reliable, and even Anthropic's account of its struggles confirms that multi-agent systems carry heavy complexity.