Why the tasks dictate the architecture

McGuinness then explains why the two teams reasonably chose differently, using the read-versus-write distinction from Chapter 1.

Anthropic's research agent does a "read" task: it collects and synthesizes information into a report. Read tasks split into independent subtopics, which makes them ideal for an orchestrator-worker design with parallel subagents.

Cognition's Devin does a "write" task: modifying code. Writing requires holding the whole codebase, the prior changes, and the user's directives in one unified context, so that no detail is neglected. Splitting that across parallel agents invites exactly the conflicting-decision failures Cognition warns about.

His conclusion: multi-agent is not a universally superior architecture, it is one that suits research-style workflows that decompose into independent, parallelizable subtasks. For inherently sequential problems, a single capable agent with unified context provides simplicity and reliability. The distinction between Devin and the Research agent is a matter of orchestration, not a fundamental technology difference. Both are built from the same parts: reasoning models, prompts, tools, and data plus memory. They differ only in how they break down and coordinate the work.