Anthropic says it’s time to stop using subagents

What Anthropic said

On October 22, 2026, Anthropic published Building agents: When to fold subagents into your main agent loop, by Diane Sarkis. It’s written for companies building agents on Claude Platform, Anthropic’s platform for developers, and its advice fits in one line: use “fewer subagents in favor of progressively disclosed context and tools.”

The finding behind it comes from Anthropic’s Applied AI team: “In our comparisons across several enterprise deployments, a single agent with skills frequently outperformed both the one-prompt-for-everything design and the subagent design on quality, and often at a lower cost and latency per task.”

The article’s examples:

• One company found that the dozen specialist subagents in its customer-facing agent made “a consistent on-brand experience difficult to maintain” and added latency that hurt the experience.
• One customer that tested its multi-agent harness against a single agent loop “saw a roughly 2x faster runtime with the latter.”
• Many proposed designs pile up an intent classifier, a fixed reasoning pipeline and a multi-agent orchestrator on top. Anthropic’s feedback is often that “one clear system prompt, 4-5 thin data tools, and firm output guardrails would cover most of the use case.”

It also names the pattern it sees most: companies giving each team or problem type its own subagent, “with the agent architecture becoming a copy of their org chart (Conway’s law at work).”

The alternative is progressive disclosure: “the agent sees short descriptions of the skills and tools, which are fully loaded only when needed.” Anthropic says it cut Claude Code’s own system prompt for its more advanced models “by over 80%,” partly by loading skills and tools that way.

Why it spread

On October 24, Tak (@cherry_mx_reds) summed up the article on X: “Anthropic is now recommending that companies REMOVE unnecessary subagents from their AI systems.” By the morning of October 25, the post had about 660,000 views and 4,400 bookmarks. Its last line: “improving agent architecture is becoming less about what you can add and more about what you can delete.”

The word that matters is “unnecessary.” Anthropic’s article says “subagents still have their place,” and that Anthropic uses them in its own agents, including its multi-agent research system. It doesn’t say to delete them all. It says to measure, then fold in the ones that don’t earn their keep.

Skills vs subagents, side by side

Both let an agent handle more than its system prompt describes. They differ in where the work happens.

	A skill	A subagent
What it is	Instructions for a kind of task, with optional reference files	A separate agent with its own context window, tools and instructions
How it loads	A short description is always in context; the rest loads when it’s used	A fresh context window each time it’s called
What it sees	The whole conversation	Only the task it’s handed
What you get back	The same agent does the work, with everything it already knows	A summary of the work
Best for	How you do something: steps, rules, judgment calls	Parallel, self-contained work, and long output you don’t need to keep

Claude Code’s skills documentation says “skill descriptions are loaded into context so Claude knows what’s available, but full skill content only loads when invoked.” Its subagent guide says each subagent “starts with a fresh, isolated context window. It doesn’t see your conversation history, the skills you’ve already invoked, or the files Claude has already read.”

Anthropic’s article draws the lines this way. Skills are for when “the agent needs to know how a team does work.” Tools are for when it “needs to take an action, read live data from a system, or return a result in a fixed shape.” Knowledge bases are for “reference material the agent only needs some of the time.”

Its main argument for one agent is shared context: “An agent that has seen the result of each earlier step can connect them, which leads to richer outputs and better insights than a set of summaries handed back by subagents.” Split the work by problem type instead, and the agents “engage in a ‘telephone game,’ passing information back and forth with each handoff degrading fidelity.”

When a subagent is still the right call

Anthropic lists three cases:

1. A single agent loop “hits measurable performance degradation from context bloat.”
2. Several tasks “need to be tackled in parallel with fresh context windows that can be discarded once the task is done.”
3. A single agent “struggles to select appropriate tools or there are conflicting instructions in the system prompt.” Tool choice, it says, “tends to degrade” once a system prompt holds more than 30 to 50 tool definitions.

Claude Code’s subagent guide draws a similar line. Stay in the main conversation when a task needs back-and-forth, when “multiple phases share significant context, such as planning, implementation, and testing,” or when “latency matters.” Use a subagent when the task “produces verbose output you don’t need in your main context,” needs its own tool restrictions, or “is self-contained and can return a summary.”

When you do split work up, Anthropic says to divide it by context, not by problem type: give each subagent “a piece of work that needs little from the rest of the system and have it hand back a compact result.” Its examples are independent research paths, separate components with clean interfaces, and black-box verification.

What subagents cost on a plan

Anthropic’s article counts cost per task, for companies paying API prices. On a Claude Pro or Max plan, the same cost comes out of your limits.

Claude Code’s subagent guide says a subagent “sends its own requests, which count toward the same usage limits as your main conversation.” Because it starts without your conversation, it may read files the main agent has already read. By default, “a subagent can spawn subagents of its own, up to three layers below the main conversation.” The cost guide lists subagents among the reasons a long session can use “far more of your plan limits than your activity suggests,” and says agent teams use “approximately 7x more tokens than standard sessions when teammates run in plan mode.”

Anthropic’s research system, which it wrote about in June 2025, shows the scale: agents “typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats.” For broad research the extra tokens paid off, but Anthropic noted that “most coding tasks involve fewer truly parallelizable tasks than research.”

A skill costs little until it’s needed. Claude Code’s documentation says “a skill’s body loads only when it’s used, so long reference material costs almost nothing until you need it.” Two catches: every skill’s description is in context on every turn, used or not, and once a skill loads, “every line is a recurring token cost.” Prune the skills you don’t use, and keep each SKILL.md under the 500 lines the documentation suggests.

Subagents can also save usage. Test runs and log files can fill the main conversation, which is sent again with every message after them. Claude Code’s cost guide suggests handing that kind of work to a subagent, so “only a summary returns to your main conversation,” and running it on a smaller model.

To see where your usage goes, run /usage in Claude Code. On Pro, Max, Team and Enterprise plans, it attributes recent usage to “skills, subagents, plugins, and individual MCP servers,” each as a share of the total.

How to fold a subagent into a skill

Anthropic’s closing advice: “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.” In Claude Code, that comes down to six steps:

1. List your subagents. Claude Code keeps custom subagents in .claude/agents in the project, and in ~/.claude/agents for all your projects. /usage shows which ones use the most.
2. Start with the ones that share context with the main work. Anthropic suggests merging specialists “that chain on each other’s output, touch the same data, or are often called together.”
3. Move the instructions into a skill. The subagent’s prompt becomes a skill in .claude/skills, with a description that says when to use it. Long reference material goes in separate files the skill points to.
4. Write down about 10 tasks you care about. Anthropic says “even a set of 10 golden examples is enough to start.”
5. Run them both ways, more than once. Look at which tasks changed, not only the overall result. In Anthropic’s test on a sample support agent, folding in a subagent lowered latency but broke a process that had been passing, and the break “only showed up in one run out of two.”
6. Keep the subagent if the skill does worse.

Teams building on Anthropic’s API can use its migration kit, which chains Anthropic’s Claude API skills for building evals, auditing prompts, improving against an eval and cutting costs into one guided workflow. It includes a demo to try on a mock system first.

One agent that knows how you work

Every subagent draws on the same plan as the conversation that started it. Most repeat work doesn’t need a team of agents. It needs one agent that knows how you do it, and a saved routine to run it the same way each time.

Enjoy runs Claude Code, Codex and the other labs’ agents on the plan you already have, and it’s built around that kind of work:

• Run a skill from the message box. Type / to choose one of Claude Code’s, Codex’s or Muse Code’s skills, from the project, your own or a plugin, and the agent runs it with your message as its instructions. A skill you just added is listed the next time you type /. See Run one of your agent’s skills.
• Save the routine as a workflow. When a task goes well, ask the agent to save it as a workflow. Each run starts one new conversation that follows the saved instructions, and you can schedule it, say every weekday, to run while your computer is awake with Enjoy open. See Turn repeat work into workflows.
• See when the agent hands work off. For Claude Code, Codex, Antigravity and Muse Code, the activity panel lists Main agent and Subagents, with how many are working. In a Claude Code conversation on your computer, Usage in its ⋯ menu shows what the conversation used, and the agent settings show how much of each limit is left. See Follow what your agent is doing and Make your agent usage last.

Enjoy is free, and the limits stay your plan’s.