A team proposes replacing its working single-agent coding assistant with a multi-agent system: a planner, a coder, a tester and a reviewer, each with its own context. The assistant makes coordinated edits across tightly coupled modules where each change depends on decisions made in the others. The budget is fixed and the task value per request is modest. Anthropic's published experience with multi-agent systems says this decision should rest on which assessment?