Anthropic's Applied AI team found one agent with skills often beat subagent setups in enterprise tests
One customer saw about 2x faster runtime with a single loop, a figure Anthropic reports from that customer alone, not from a broad study.
In brief
Anthropic: one agent often beats many
Anthropic's Applied AI team says a single agent with skills often beat subagent designs. One customer saw about 2x faster runtime. Subagents still fit some jobs.
Anthropic's Applied AI team says that in several enterprise deployments, a single AI agent with added "skills" often did better than designs built from a router and many subagents. The claim comes from guidance dated 22 October 2026, published on claude.com.
A subagent is a helper agent with its own memory, tools and instructions, directed by a coordinator. Think of an office where every request passes through a receptionist and then a dozen specialists.
What Anthropic reports
Across those comparisons, the single agent with skills frequently beat both a one-giant-prompt design and the subagent design on quality. It was often cheaper and faster per task too.
One company found that its dozen specialist subagents made a consistent on-brand voice hard to keep and added delay that hurt users.
One customer compared its multi-agent setup with a single loop and saw roughly 2x faster runtime. That is one customer's result, reported by the vendor. It is not a general study.
Why it matters
Many companies split an agent by team, so the software copies the org chart. Anthropic points to Conway's law, the idea that systems mirror the organisations that build them. It says reviewed designs often stacked a classifier, a fixed reasoning pipeline and an orchestrator, when a single clear prompt, a handful of slim data tools (four or five) and strict limits on outputs would meet most needs.
A single loop can also load skills only when needed, instead of carrying everything at once.
Subagents are not dead
Anthropic says subagents still fit in three cases: context bloat that measurably hurts performance, parallel tasks that need fresh context windows, and trouble choosing tools or conflicting instructions. It uses them in its own multi-agent research system.
One outlet, HuggingNews, framed this as advice to cut subagents. That overstates Anthropic's wording, which is conditional.
The companion kit
Anthropic also released single-agent-loop-kit on GitHub under Apache-2.0. It helps teams measure a system, merge subagents that always run together, and fold the rest into skills one at a time, keeping only changes that score as well on the same test.
Some reports call it a Claude Platform migration tool. The repo itself says it is a reference implementation, not maintained and not accepting contributions.
The evidence is Anthropic's own and comes from unnamed enterprise customers. No independent test was found, and the 2x figure is vendor-reported.