Anthropic offers a kit for reassessing subagents
Its guidance makes consolidation a decision to evaluate.
Anthropic’s guide proposes moving some specialist agents into a shared working conversation. Procedures and tools load when needed, allowing the main agent to retain earlier context. This is a choice for developers reviewing an existing system, not a requirement to remove every specialist. The company still identifies independent verification, parallel tasks with separate context and an overloaded main context as reasons to retain subagents.
The kit audits the system first, then compares evaluations around individual changes; a person decides whether each stays. It pauses before paid runs and provides a token estimate. Its offline support example needs no API key, but uses keyword rules and checks the plumbing without scoring quality. This is an unmaintained reference implementation, not a validated migration for arbitrary systems. Teams need a safe test route outside production and someone who can define a correct outcome.
Its commerce guidance explains why the boundary matters: a returns conversation may need order history, the current cart and the catalog together. A narrow research task, by contrast, can justify its own context. The separation should therefore reflect the information the work needs.
The Decoder also covered the recommendation, emphasizing that teams should evaluate consolidation instead of merging blindly.