- Anthropic (22 Oct) advises replacing many subagents with one loop that loads skills when needed.
- At one customer the single loop ran about 2x faster.
- The evidence is mostly anecdotal: change one piece at a time and measure.
In short: in an article dated 22 October 2026 (author: Diane Sarkis), Anthropic argues that in many companies one agent that loads skills and tools only when needed works better than a team of specialised subagents. Teams often build agents that mirror their own org chart, and that causes problems.
What Anthropic proposes
- One model and one context window in a tool loop, with turn and cost limits set in the system around the model.
- Domain procedures in skills, and hard rules such as validation and approvals in tools.
- Subagents stay where context really bloats, tasks run in parallel, or one agent gets lost choosing tools.
- According to the article, tool choice starts to degrade beyond roughly 30–50 definitions in the system prompt.
Numbers from the article
- At one customer the single loop ran about 2x faster than a setup with many agents.
- Anthropic cut the Claude Code system prompt by more than 80% for more advanced models, partly by loading skills and tools progressively.
- In a sample support agent, folding one subagent into the main one cut latency but broke a process. The failure showed up in only one of two runs, which is why tests should be repeated.
How to measure it
The article recommends starting with a task-completion test (even about 10 golden examples), folding subagents in one at a time and comparing results with and without each. Anthropic also published the repository anthropics/single-agent-loop-kit (Apache-2.0 licence). According to its README it is a reference implementation that is not maintained.
How to read this
Most of the evidence is anecdotal: comparisons from a few deployments and one latency figure, with no description of the method. Treat it as a hypothesis to test on your own system, not a rule. I write more about building and testing such systems in the harness and evals hub.