The Pulse
Anthropic Says Single-Agent Loops Can Beat Subagent Systems
Anthropic says some enterprise agents perform better when specialist subagents move into a single agent loop with skills and tools. Its new migration kit asks teams to compare each change against task-completion evaluations and keep subagen
AI.info Team · Written with AI · fact-checked ·
“As models get more capable and agents get longer running, more of your subagents can be skills and tools.”
Diane Sarkis, author of Anthropic’s article
Anthropic published that recommendation on October 22, alongside an open-source kit for moving some multi-agent systems toward a single agent loop. The company says its Applied AI team has seen single agents equipped with skills outperform both sprawling prompts and subagent designs on quality across several enterprise deployments, often with lower cost and latency. The claim is a case for testing simpler designs—not a finding that one agent always beats a team.
Anthropic’s Case for Fewer Subagents
Anthropic describes a common pattern in large organizations: teams build separate agents for separate capabilities, then add a coordinator to route work among them. The resulting systems can lose information in handoffs, accumulate overlapping instructions and slow down as agents exchange summaries. A customer told Anthropic that about a dozen specialist agents in its customer-facing system made a consistent brand voice harder to maintain.
The company says a single agent can instead load skills and tools as needed, keeping relevant work in one context without placing every instruction in its prompt at once. Anthropic also says it cut Claude Code’s system prompt for more advanced models by more than 80%, including changes tied to this progressive loading approach. It does not publish detailed task counts or comparison tables for the enterprise deployments described in the article.
One Customer’s Harness Ran Roughly Twice as Slowly
Anthropic offers one specific runtime comparison: a customer that tested its multi-agent harness against a single-agent loop saw the latter run roughly twice as fast. The article gives no matching cost figure for that customer, so the speed result should not be read as a quantified cost saving. Anthropic’s broader claim is that the single-agent design often lowers both cost and latency per task, but the post does not break out results by deployment.
The company also draws boundaries around the advice. It says subagents can still help when a task needs independent work in parallel, when a single agent’s context grows enough to hurt performance, or when a model struggles to choose among tools or reconcile conflicting instructions. Anthropic’s own multi-agent research system is one example it cites for keeping agents in parallel work.
The Migration Kit Makes Teams Test Each Change
The single-agent-loop-kit is a Claude Code skill and reference project, not a packaged claim that every system should be consolidated. It audits existing subagents, recommends whether to merge, fold, keep or call each as a tool, then compares changes against the same evaluation set. Teams can start without an existing eval: the kit helps draft test tasks, but asks a person to define what counts as a correct result.
Anthropic’s process calls for changing one part at a time and asking before paid evaluation runs. The repository includes scripts to compare task completion, latency, model calls, tokens and cost, along with a sample support agent that starts with six subagents. Its skill says plainly that keeping a subagent is a valid result when the evidence supports it.
The Published Demo Has a Narrow Test Record
The repository’s own account limits how much the kit has been tested: it reports two real-model runs of its example, conducted by one person on one model. The runs covered a 12-task baseline and a 17-task baseline, with one subagent folded in one run and another called as a tool in the other. Anthropic says the merge and upkeep steps, the finished single-loop example and use on systems beyond the sample have not been tested on a real model.
That caveat matters because the recommendation depends on the work being measured. A single loop may avoid coordination overhead, but folding a specialist into it can also remove a behavior the existing system gets right. Anthropic’s kit is built around that tradeoff: compare the revised system on representative tasks, inspect which outcomes changed, and remove a subagent only when the replacement holds up.