Anthropic Says One Agent Loop Ran 2x Faster Than Subagents

Anthropic is telling companies that build AI agents on its platform to consider using fewer subagents and to fold specialist agents back into one main agent loop, in an article on claude.com dated 22 October 2026 by Diane Sarkis. “One customer that evaluated their multi-agent harness against a single agent loop saw a roughly 2x faster runtime with the latter,” the article says.

Anthropic says its Applied AI team found a pattern of enterprise deployments that benefited from folding routers, planners and domain subagents into the main loop. In its comparisons, it says, a single agent with skills frequently outperformed both a one-prompt design and a subagent design on quality, often at lower cost and latency per task.

It also published a migration kit on GitHub to help teams test the change, and warns that folding can break tasks that used to pass.

Business Pill · THE MODEL AND THE HARNESS

A short Business Engineer explainer on the difference between an AI model and the harness around it. It teaches the general idea only and says nothing about Anthropic’s article.

What Anthropic Recommends

The article’s complaint is about structure that copies the company rather than the task. It says enterprises often give each team or problem type its own subagent, “with the agent architecture becoming a copy of their org chart (Conway’s law at work).”

It lists the costs it sees in large organisations: details lost in handoffs between agents, “hundreds of overlapping subagents with no enforced standards and no specific owner”, a system prompt where every team adds a sentence, evals tuned to each team’s subtask, and added latency.

Its suggested design is one model and one context window running a tool loop, with turn and cost budgets set around it. Team procedures move into skills, which the agent loads only when needed, and actions and live data move into tools. Anthropic says it reduced Claude Code’s system prompt for more advanced models by over 80%, including cuts from loading skills and tools this way.

The article keeps a place for subagents. It says to use them when one loop shows measurable degradation from context bloat, when tasks need to run in parallel with fresh context windows, or when one agent struggles to pick the right tools or faces conflicting instructions.

What Was Posted on X

A post in Chinese on 24 October 2026 by the @Xudong07452910 account summarising Anthropic’s article, including the roughly 2x faster single loop, the support-agent case that broke after folding, and the migration kit. We use it as the relay that pointed us to the article; every figure in this piece comes from Anthropic’s article itself.

Subagents vs One Loop, in Plain Words

A subagent is a separate AI worker with its own instructions, tools and working memory, called on by a coordinator that hands it a task and gets a summary back. Anthropic says that when agents are split by problem type, they play a “telephone game,” with each handoff degrading fidelity.

The single-loop design keeps one worker in charge of the whole conversation. It sees short descriptions of every skill and tool and opens the full instructions only when a task calls for them, so the context it has built up stays in one place.

The Figures in Anthropic’s Article

Each row is a figure the article gives. They describe different things, so we list them and do not combine them.

Where Folding Went Wrong

The article reports its own failure case. When Anthropic tested a sample support agent, “folding one subagent into another lowered latency, but also broke a process that had been passing.” The main agent held the order lookup tool and answered on its own, without asking the shipping specialist for the delivery date.

Its advice is to check which tasks flipped, not only the final score, and to run each task more than once, “since a break like this one only showed up in one run out of two.” Its closing step-by-step: “Build a small end-to-end eval, fold in one subagent, and keep it as a subagent only if the folded version scores worse, then rinse and repeat.”

The migration kit sits in a public GitHub repository, anthropics/single-agent-loop-kit, created on 15 October 2026 under the Apache 2.0 licence, according to GitHub. The article says it runs a guided workflow and includes a demo to try on a mock system first.

Connecting the Coverage

Skills, the building block the article leans on, have been moving fast. On 22 October Anthropic said it was donating Agent Skills to the Agentic AI Foundation, with 1 million skills published and nearly 280 million installed, as we reported in Anthropic Donates Agent Skills After Nearly 280M Installs.

The harness around a model has been where much of this month’s agent news sits. A Google and Northwestern paper defined the harness as “the prompts, memory, skills, tools, and orchestration through which the model performs tasks” and reported gains of up to 11.8 points from editing it, as we covered in Google’s Harness² Improves AI Agents by Up to 11.8 Points.

On cost, Google said its Agent Substrate runtime delivers as much as 10x higher compute density for agents, in Google Says It Can Run Up to 10x More AI Agents per Machine. And on 15 October the Claude Code team said Claude writes about 90 percent of Claude Code’s own code, in Claude Writes 90% of Claude Code as Reverts Fall Over 40%.

What’s new here: Anthropic’s own guidance to enterprises to cut the number of subagents, with a runtime figure from one customer and a failure case from its own test.

The Structural Read

As we read it, the article’s case rests on three stated reasons: shared context in one window, progressive disclosure of skills and tools to avoid context bloat, and newer models that can manage more in a single loop.

As we read it, the evidence offered is Anthropic’s account: one customer’s roughly 2x faster runtime, one company with a dozen specialist subagents, and a Claude Code system prompt cut by over 80% for more advanced models.

As we read it, the article pairs the advice with its own warning: in its sample support agent, folding one subagent lowered latency but broke a process that had been passing, and the break showed up in one run out of two.

Three Implications

FOR TEAMS RUNNING MANY SUBAGENTS The article says to track which tasks flipped after each fold, not only the final score, and to run each task more than once.

FOR AGENT BUILDERS ADDING TOOLS Anthropic recommends tools grouped by user intent, saying tool choice tends to degrade once a system prompt holds more than 30-50 tool definitions.

FOR TEAMS THAT KEEP SUBAGENTS The article keeps them for context bloat, parallel tasks with fresh context windows, and agents that struggle to pick tools or face conflicting instructions.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework puts it as “The surface, not the model, governs autonomy. The agentic surface — the harness — is becoming where value pools in the entire AI stack.”

As we read it, Anthropic’s article is a design note on that harness: how many agents it holds, and how skills and tools are loaded into one loop. The framework was published in July 2026 and does not discuss this article.

What Is Not Established

The article does not name the customer behind the roughly 2x runtime figure or the company with a dozen subagents, and it gives no cost figures for the comparisons. The results it describes are Anthropic’s account of customer deployments and of its own tests; we found no independent measurement in the texts we read.

The Bottom Line

Anthropic’s 22 October article tells enterprises to fold subagents into one agent loop one at a time, keeping each only if the folded version scores worse, and says one customer saw a roughly 2x faster runtime with a single loop.