Claude Sub-Agents & Managed Agents, Explained Simply

Claude sub-agents are copies of the main agent that work on separate parts of a task at the same time, each in its own space, then report back so the main agent can combine the results. They're how Claude tackles big jobs faster and more reliably than working through everything in a single thread. This guide explains what sub-agents are, how the agent loop behind them works, when Claude uses them, and how they differ from the separate product called Claude Managed Agents.

You don't need to set any of this up by hand. It's worth understanding because it explains why some results are noticeably better, and how to ask for them.

Watch: our Claude Managed Agents review walks through the managed-agents side in detail.

What are Claude sub-agents?

A sub-agent is a parallel instance of Claude Cowork that the main agent creates to handle one slice of a larger task. Imagine asking Claude to research five competitors. Without sub-agents, it works through them one after another, and quality tends to drift after the first one or two as the context fills up. With sub-agents, it spins up several versions of itself, one per competitor, all working at once.

The results are better, faster, and less biased, because the agent researching competitor two isn't coloured by what it found for competitor one. Each sub-agent works in its own context window, does its slice, and reports a summary back to the main agent, which then synthesises everything into the final output.

How the agent loop works

Sub-agents make more sense once you understand the loop they sit inside. When you give Claude a task, it doesn't just answer. It makes a plan, takes an action, checks its own work, and then decides what to do next, repeating until the job is done.

That self-checking loop is the core difference between an agent and a chatbot. A chatbot replies and stops; an agent keeps going, correcting course as it learns from each step. Sub-agents are simply that loop, run several times in parallel, with a lead agent holding the overall plan and stitching the pieces together. Anthropic's own engineering team describes this as an orchestrator-workers pattern, where a central agent breaks down a task, hands pieces to workers, and combines what comes back.

When does Claude use sub-agents?

Most of the time Claude decides for itself. A task like "research these five companies and compare them" is usually enough to trigger parallel agents, because the work splits cleanly into independent pieces.

You can also nudge it. Adding something like "run these in parallel" or "use one sub-agent per item" to your prompt makes the behaviour explicit. The thing to remember is that the main agent only sees each sub-agent's final summary, not every step it took, so the instructions you give matter: each sub-agent needs enough direction to produce a useful result on its own.

Sub-agents inside skills

Sub-agents aren't only an ad-hoc trick; they can be built into a skill so a repeatable workflow uses them every time. Our own video-highlights skill is a good example. After it reads a transcript and picks candidate clips, it hands off to a dedicated visual-analyzer agent that screenshots each candidate so the final choice accounts for what's on screen, not just what was said.

That's the pattern worth copying: when a job has a step that benefits from focused, parallel attention, a skill can spin up a specialist agent for exactly that step. When you build skills with the Skill Creator, you can simply ask whether sub-agents would help, and it sets them up.

Sub-agents vs Claude Managed Agents

The names are similar, so it's worth being clear. Sub-agents are the parallel helpers described above, created on the fly inside a task. Claude Managed Agents is a separate, developer-facing product: a managed harness and infrastructure for running Claude as an autonomous agent on long-running, asynchronous work, without building your own agent loop and sandbox.

For a business user in Cowork, sub-agents are the relevant idea, they happen automatically and make your results better. Managed Agents matters when a developer needs to run agents at scale in their own systems. Both lean on the same principle, that breaking work across focused agents beats forcing everything through one.

A real sub-agent in action

Our video-highlights skill is the clearest example we have. After the main agent reads a workshop transcript and shortlists eight to ten candidate clips, it can't tell from the words alone which ones will work as standalone videos, because that depends on what's on screen. So it hands each candidate to a dedicated visual-analyzer sub-agent, running in parallel, that looks at screenshots from the recording and judges whether the speaker is showing a meaningful slide, doing a live demo, or just sitting in gallery view. Each sub-agent reports a verdict, the main agent picks the final five, and a separate clip-cutter agent does the editing. One job, split across focused workers, each seeing only its own slice.

That's the pattern worth stealing: when a task has a step that benefits from focused, parallel attention, a sub-agent handles exactly that step without cluttering the main agent's context.

How to ask for sub-agents

Most of the time you don't have to. A prompt like "research these five competitors and compare them" usually triggers parallel agents on its own, because the work splits cleanly. When you want to be explicit, add a line to your prompt: "run these in parallel" or "use one sub-agent per item." And when you're building a skill, the Skill Creator will ask whether sub-agents would help and set them up for you.

One thing to keep in mind: the main agent only sees each sub-agent's final summary, not every step it took. So each sub-agent needs enough direction in the brief to produce a useful result on its own, the same way you'd write a clear task for a colleague you won't be checking in with mid-way.

Common questions about sub-agents

Do sub-agents cost more usage? They do more work in parallel, so a sub-agent task uses more than a single-thread one, but it's often faster and the quality is higher because each agent stays focused.

Can I see what each sub-agent did? You see the main agent's synthesis and its plan. The individual steps happen in each sub-agent's own context, which is exactly what keeps the main thread clean.

Are sub-agents the same as Claude Managed Agents? No. Sub-agents are spun up on the fly inside a task; Managed Agents is a separate developer product for running agents at scale.

When sub-agents help, and when they don't

Sub-agents shine when a task splits into independent pieces that can be worked at the same time, researching several companies, analysing many files, or assessing a batch of candidates where each item stands alone. The parallel structure is faster and, because each agent starts fresh, the quality stays high instead of drifting after the first one or two.

They add little when a task is strictly sequential, where each step depends on the last, since there's nothing to parallelise. Forcing sub-agents onto that kind of work just adds coordination for no gain. The judgement call is simple: if you could hand the pieces to several colleagues to do at once, sub-agents fit; if one person has to do step two before step three, a single agent is the right tool.

What this means for your work

You don't have to manage any of this, but knowing it exists changes how you brief Claude. When a task naturally splits into parts, say so, and let it parallelise. When you find a workflow that benefits from a focused sub-step, build it into a skill. And when a developer on your team talks about Managed Agents, you'll know it's the infrastructure version of the same idea, not a different concept to learn from scratch.

If you want your team building agentic workflows like these around your own processes, our hands-on AI automation training takes operators from their first task to automations that run across your tools.