Anthropic Urges Teams to Fold Subagents Back In: A Single Loop With Skills Beats Orchestrators

Anthropic's Applied AI team has published a post on when subagents should be folded back into the main agent loop. The conclusion, drawn from real enterprise deployments, is an uncomfortable one for anyone who spent the past year building orchestrators: a single agent with skills often beats both the "one prompt for everything" setup and a sprawl of subagents on quality. It also comes out cheaper and faster. At one customer, the single loop ran about 2 times faster than the multi-agent scaffolding. At another, a dozen specialist subagents inside a customer-facing agent broke the brand's consistent tone and added latency.

Conway's Law for Agents

The authors describe a familiar organizational dynamic. Each team builds its own subagent, and a platform team puts an orchestrator on top. The result is an agent architecture that repeats the company's org structure, along with all its ailments:

• context is lost in handoffs between agents;
• hundreds of overlapping subagents appear, with no standards and no owner;
• a "Frankenstein prompt" emerges, to which every team has added a sentence;
• decisions are made on local evals that do not reflect the quality of the system as a whole;
• latency grows, because every hop between agents is one more model call.

In terms of token economics, this is a double hit. Every task handoff requires re-explaining the context, which means extra input tokens and a risk that a needed detail gets lost. The "telephone game" effect hurts quality, and the extra rounds of calls hurt response time.

One Loop, One Window, Progressive Disclosure

Anthropic's proposed alternative looks like this:

• one loop and one context window, with limits on the number of turns and on budget;
• progressive disclosure: the model sees short descriptions of skills and tools and loads the full versions only when it needs them. This is how the team cut the Claude Code system prompt by more than 80%;
• tools are grouped by user intent, not by API. After 30–50 tool descriptions, selection quality starts to degrade;
• access rights are checked inside the tools, and the model receives a clear refusal. That way a single loop can serve users with different access levels, and there is no need for separate agents for each role.

The last point is especially practical. Instead of multiplying agents for different user groups, authorization moves to the tool level, and the agent learns to respond sensibly to a refusal.

When Subagents Are Still Needed

The authors do not call for abandoning subagents entirely. They name three justified cases:

• there is measurable degradation from a bloated context;
• you need parallel tasks with a fresh context;
• the agent gets confused by its tools and contradictory instructions.

Even then, they recommend splitting subagents by context, not by task type. Splitting by task type is exactly what produces the telephone game.

Migration Plan

The post lays out the steps as follows:

1. Collect examples. Even 10 is enough; they will serve as the basis for comparison. 2. Merge the most heavily used and tightly coupled subagents. 3. Call the remaining ones as tools. 4. Convert them into skills one at a time, comparing the result against the reference after each step.

Ownership changes organizationally too. Teams stop owning agents and start owning skills and tools. One small central team is responsible for the system prompt and for overall quality. For the migration, Anthropic has published a skill migration kit on GitHub.

What to Do Right Now

If you run a multi-agent setup in production, start with an audit: count your subagents, measure the average latency at handoffs, and check the size of the system prompt. Put together a set of 10 real cases and run it through a single loop with skills. If quality does not drop while time and token spend fall, the setup can be simplified. Separately, check whether your toolset exceeds 30–50 descriptions. If it does, group the tools by user intent. And name an owner for the shared prompt and evals up front, or the "Frankenstein" will come back, this time in the form of skills.

According to the author of the review, OpenAI reached a similar conclusion in its post on Forward Deployed Engineers. The trend is clear: a year ago orchestrators were a showcase of maturity, and now they increasingly look like legacy.