Anthropic's advice for agent builders: fold most subagents back into one loop with skills

The guide, from Anthropic's Applied AI team, describes a familiar setup: each team builds its own subagent, a platform team puts an orchestrator on top, and the system prompt collects a sentence from every team. The problems it lists are lossy handoffs between agents, overlapping subagents nobody owns, conflicting rules and extra latency. One customer saw roughly 2x faster runs with a single loop, per the post.

The design it suggests is one model and one context window running a tool loop, with turn and cost budgets set in the harness. Team procedures go into skills, actions and live data go into MCP tools, and tool search loads definitions only when needed. Anthropic says tool choice tends to get worse past 30 to 50 tool definitions, so it recommends grouping tools by user intent instead of one tool per API.

Subagents still make sense in three cases, according to the guide: measurable context bloat in one loop, parallel tasks that need fresh context you can throw away, or a single agent that keeps picking the wrong tool.

The migration advice is gradual. Build a small end-to-end eval (10 golden examples is enough to start), fold in one subagent at a time, and look at which tasks flipped, not only the total score. In Anthropic's own test on a sample support agent, folding one subagent cut latency but broke a process that had passed before, and the break showed up in only one run out of two.