Claude Managed Agents and Dispatch Explained
What Claude Managed Agents and Dispatch are, how they differ, what runs cost and which limits apply. Checked against Anthropic docs on 19-10-2026.
Published 19-10-2026. Status and prices last verified on 19-10-2026, reviewed monthly because both products are changing.
Quick answer
This page is for the engineering lead or operations owner who wants an agent to run for hours on a task, such as a research job or a recurring report, without building the containers and the loop around it. Claude Managed Agents is Anthropic's hosted agent harness on the Claude Platform: you define the agent and an environment, Anthropic runs the loop in a sandbox. It is in beta, and it adds $0.08 per session-hour of active runtime on top of token costs. Dispatch is a different product: a long-running agent in Claude Cowork that you message from your phone and that works on your desktop. People mix the two up, and the documentation we read does not connect them.
How we checked the information in this article
We read the Managed Agents overview and quickstart, the Anthropic announcement, the pricing page, the Dispatch documentation and the Dispatch help center article on 19-10-2026. We have no console session with Managed Agents of our own, so we ran no agent and measured no cost per run, no median duration and no concurrency limit. Where the documentation is silent, this page says so.
Our gate: a capability goes in only if an Anthropic page we could open on 19-10-2026 states it, and anything we could not confirm is written as not documented. The gate rejected three things. We dropped the "10x faster" headline numbers, because the announcement bases them on customer quotes and gives one measured figure only: up to 10 points higher task success than a standard prompting loop in an internal test on structured file generation, which is Anthropic's number and not ours. We did not describe Dispatch as part of Managed Agents, because neither page links them. And we used Anthropic's own pages for Dispatch and left out third-party blog descriptions.
What are Claude Managed Agents?
Anthropic calls it a pre-built, configurable agent harness that runs in managed infrastructure. You do not build the agent loop, the sandbox or the tool execution. It is built for long-running and asynchronous work, where the Messages API is built for custom loops with fine control.
Four concepts carry the documentation. An agent is the model, system prompt, tools, MCP servers and skills. An environment is where sessions run: an Anthropic-managed cloud sandbox or a self-hosted sandbox on your own infrastructure. A session is a running agent instance doing one task, and events are the messages between your application and the agent. The built-in tools are bash, file operations, web search and fetch (optionally limited to allowed or blocked domains) and MCP servers.
Status matters here. The documentation says Managed Agents is in beta and that behavior may change between releases. Multi-agent coordination and outcome-based work are research previews you request access to, and MCP tunnels and dreaming are in a more limited preview. Every request needs the managed-agents-2026-04-01 beta header, which the SDKs set for you. The quickstart creates an agent with the ant command line tool (ant apply on a markdown file with a model and tool list) and a session through the Python SDK. We did not run it.
What does Dispatch do?
Dispatch lives in Claude Cowork, not in the Claude Platform API. The documentation describes a long-running agent that takes a high-level instruction in one conversation, splits it into child tasks and runs each as a separate Cowork or Code session. Coding work goes to Code against a workspace you already set up, and knowledge work goes to Cowork in the project you name. Child tasks do not spawn children of their own.
From your phone, you message Dispatch and the work runs on your desktop, which must be awake with the app open. A permission prompt that you do not answer within ten minutes is denied automatically. The help center article says Dispatch is in limited beta for some Pro and Max plans and is not available to new users, so check whether you see it in your sidebar before you plan around it. The two pages also disagree on platforms: the documentation names macOS and Windows, the help center article adds Linux and Windows x64. We would trust the help center article, which was updated more recently, and tell you to check your own app.
If you search for Dispatch to run unattended server jobs, it is the wrong tool. It is one thread for one person, on one computer, and our productivity skills are the collection for that kind of personal task handling. Our guide to background agents with skills and to scheduled tasks cover the options that do not depend on your laptop.
Managed Agents, subagents, agent teams or Dispatch: which one?
Choose by who hosts the agent and who starts it. This is our reading of the Anthropic pages, not a measured comparison.
| Option | Who hosts it | Who starts it | Our verdict |
|---|---|---|---|
| Managed Agents | Anthropic, in a cloud or self-hosted sandbox | Your application, through the API | Where we would start for a product feature that runs for hours |
| Agent SDK | You, one subprocess per session | Your application | We choose it when you need the loop in your own process |
| Subagents and agent teams | Inside one Claude Code session | The main agent | Fine for splitting one job, not a deployment model |
| Dispatch | Your own desktop | You, from a phone | A personal assistant, not infrastructure |
Our guide to agent teams with skills and our overview of multi-agent orchestration go deeper on the middle two rows. Our agents and orchestration skills are the matching collection.
How is a Managed Agents run billed?
Managed Agents is priced on consumption: standard token rates plus $0.08 per session-hour for active runtime. Only time in the running state counts, and time spent idle, rescheduling or terminated does not. This replaces the container-hour charge for code execution, so you do not pay twice. All of this is vendor-stated on the pricing page and the announcement, read on 19-10-2026. Token rates per model are in our Claude API pricing guide.
The pricing page gives one worked example: an hour of coding on Claude Opus 5.5 with 50,000 input and 15,000 output tokens costs $0.58 including the $0.08 runtime. We checked the arithmetic and it holds: $0.20 of input at $4 per million tokens, $0.30 of output at $20 per million tokens and $0.08 of runtime.
| Active runtime | Runtime fee, our calculation |
|---|---|
| 1 hour | $0.08 |
| 10 hours | $0.80 |
| 100 hours | $8.00 |
Our verdict: the runtime fee is small against the tokens. A full day of active runtime is $1.92 in fees, about the price of two Opus 5.5 tasks of 100,000 input and 20,000 output tokens, which cost $0.80 each by our calculation. Plan the model and the context size first. Our workflow automation skills are where we would look to keep runs short and repeatable.
Which limits and restrictions apply?
The documentation names three. It is beta, so behavior can change. It is stateful by design, with sessions, sandbox state and outputs stored server-side, and you can delete sessions and uploaded files. And it is not eligible for Zero Data Retention or a HIPAA Business Associate Agreement. For a team in health care or one with a zero-retention contract, that last point settles the question before cost does, and our security skills include checks for what an agent may reach.
It does not publish a concurrency limit, a rate limit number or an average run time on the pages we read. Our run of the product would have produced a cost per run, a median duration and the number of concurrent sessions on one plan. We have none of those, and we would not trust a number for them from a blog post. If you need them, read the rate limits reference and test one agent on your own account. Our business operations skills are the collection for the recurring back-office jobs that usually motivate this question.
Which skills suit a managed agent?
An agent definition can carry skills, so a written procedure travels with the agent. We have not run any listing below inside Managed Agents, and we judge them only by what the listing says. The pattern we look for is a job with a clear end, no human at the keyboard and files as the output.
Deep research team runs a hierarchical team of agents for deep research and is free by its listing. Agentic workflow is a risk-aware engineering protocol for agent tasks at $5. Endless loop is an autonomous research and task loop that builds on earlier findings at $12, and it is the one we would put a budget cap around first. Agent handoff orchestrator generates structured handoff packages for multi-agent work and is free. We sell skills that load into Claude, Cursor, Codex CLI and Copilot, and we name ours here. No vendor paid to be listed.
When should you self-host instead?
Self-host with the Agent SDK when the loop must live in your own process, when you need your own session store or when your compliance team requires zero data retention. The SDK hosting guide describes the container patterns, and our Agent SDK guide covers them. Managed Agents also offers a self-hosted sandbox if you want Anthropic's loop on your own infrastructure. Our DevOps and cloud skills cover the container side.
One more term to untangle: Claude Code managed settings are an administrator feature for pushing policy to users, and have nothing to do with Managed Agents. Anthropic's managed settings documentation (listed under Sources) covers that.
Where to start
Create one agent with the ant tool, give it one narrow task and read the session events before you add tools. Then check the billing page for the run. We review this page every month and move items out of the beta caveats as Anthropic does. Written by Samuel Rose, Founder of Agensi.
Sources
All read on 19-10-2026.