Anthropic: Fold Subagents Back Into a Single Loop — One Customer's Tasks Ran Twice as Fast

On October 22, Anthropic published an article on when to fold subagents back into an agent's main loop. A subagent is a separate AI worker to which the main agent hands part of a task: finding files, checking a result, collecting data. The main loop is when a single model thinks for itself, calls tools, looks at the result and repeats the cycle until the task is done. The article's central point: at every transition between agents, the context has to be retold from scratch, and some details are lost. According to the article, one customer's tasks ran roughly twice as fast after such a merge.

Where an agent chain loses meaning along the way

Think of a game of telephone where the players are language models. The first agent receives the whole task. It passes a short summary to the second. The third gets a summary of the summary. Each step compresses the context, and nuances, constraints and small conditions fall off first.

The author of a write-up in the Telegram channel "Technokafe" explained this with a scene from "Kin-dza-dza!". The elder orders a gravitsappa, but the order travels down a chain of Patsaks. The fifth retells it to the fourth, and in the end they bring back a wheel. In the author's video, Claude dissolves such a chain, while Codex works alone and pulls the skills it needs out of a drawer when it needs them. Skills here are pre-written instructions for a specific type of work. They load at the moment they are needed, with no separate agent and no retelling of the task.

Handoffs also carry a direct cost. Every transfer is an extra request to the model, a retelling and a wait. If a chain of several agents does what one agent with full context could do, you pay in time and tokens for lost information. Apparently, that is where the customer's twofold gain came from.

What breaks after a merge and how to catch it

The article does not sell merging as a free improvement. It includes an example from customer support: the agents were merged, and one agent forgot to ask the logistics contact for the delivery date. Previously, a separate specialized agent asked that question, and its very existence guaranteed the question would be asked. After the merge, the guarantee was gone.

That is why Anthropic advises running an end-to-end check after any merge. In practice:

• Take ten reference tasks for which you know the correct result. That set is enough to start.
• Run them before the merge and after it.
• Do not look at the overall score. It can rise and hide a failure in a single scenario.
• Look for tasks that used to pass and now fail. That is usually where the lost detail sits, like the delivery-date question.

Subagents are not being abolished. The article leaves them three jobs:

• parallel reconnaissance, when several agents search for information in different places at the same time;
• independent verification, when a second agent looks at the first one's result with fresh eyes, not contaminated by its reasoning;
• context isolation, when heavy work with a large volume of data should not clutter the main conversation.

Anyone who wants to try this hands-on can open Anthropic's reference single-agent-loop-kit. It is unmaintained, meaning you should not expect updates, but it shows the single-loop pattern.

What the gravitsappa video cost

The post's author attached the economics of their own experiment to the write-up. The video was assembled entirely with AI tools:

• Script, assembly and editing in Claude Code on the Opus 5.5 model: about $7.3 at API rates. That is roughly 0.8% of the weekly limit on a Max subscription.
• The frames were drawn by GPT Image through Codex on a ChatGPT subscription, with no separate charge.
• Scene videos were made by Seedance 2.0 at 480p resolution through the kie service. Volcengine matched the lip movement to the author's voice. Together these two items cost 1 849 credits, about $9.2.
• Roughly $1 of that sum was wasted: the connection dropped, and one clip was ordered twice.
• ElevenLabs voices: about $0.08.
• The author built the 1986 film grain, credits and audio with code.

The total comes to about $16.6, not counting subscriptions. For a short video with voiceover and animation, that is a small figure. It also shows where money leaks: not on expensive models, but on repeat orders after a dropped connection. The same principle applies to agent chains: extra handoffs and repeats cost more than they seem to.

What engineers should do

Audit your own chain. For each subagent, ask: is it there for parallelism, independent verification or context isolation? If it serves none of these and only separates roles, it can probably be folded into the main loop. First assemble ten reference tasks, then merge and compare task by task, not by the average score.