Anthropic guide: fold most subagents back into one main agent loop with on-demand skills and tools; one customer ran about 2x faster

Background and the problem

Over the past year, "an orchestrator plus a set of specialist subagents" became the default enterprise agent design. In this official article, Anthropic's Applied AI team argues that many companies' agent architecture has turned into a copy of their org chart (Conway's law): each team builds its own subagent and a platform team adds an orchestrator on top. That leads to four problems: details get lost when context passes between agents; subagents sprawl into hundreds of overlapping agents with no owner; every team adds a line to the system prompt, producing a "Frankenprompt" of conflicting rules; and each team's evals target only its own subtask, so nobody owns the whole.

How it works

The article recommends one model and one context window running an open-ended tool loop: each turn the model emits a batch of tool calls, the harness runs them in parallel and returns raw results to the same model, which decides the next step, while turn limits, cost budgets and write gates live in the harness. Domain procedure and ordering move into skills; validation, approvals and memory lookups move into tools. The model sees short descriptions and loads the full content only when needed, which is progressive disclosure. Anthropic gives three reasons: shared context lets the agent connect every earlier step; progressive disclosure removes the main reason teams split work across subagents; and starting with Opus 4.5, models can manage far more in a single loop than most teams expected.

Benchmarks and measured results

The article publishes no standardized benchmark, so there are no official benchmark numbers. Three vendor-reported figures can be checked: across several enterprise deployments, a single agent with skills frequently beat both the one-prompt and subagent designs on quality; one customer saw roughly 2x faster runtime after switching to a single loop; and Claude Code's system prompt for more advanced models was cut by over 80%. The article also shares a counterexample: on a sample support agent, folding in a subagent lowered latency but broke a process that had been passing, and the break showed up in only one run out of two.

What developers can do now

If you already run a multi-agent system, the companion single-agent-loop-kit lays out three steps: build an end-to-end eval and score the current system (10 golden examples are enough to start); merge subagents that always run together or touch the same data; then fold the remaining subagents into skills plus MCP tools one at a time, comparing before and after and checking which tasks flipped. For subagents you keep, have the main agent call them as tools and hand off fully only when they must talk to the user directly. This applies whether you build on Claude Code, the Agent SDK, or your own harness.