Claude Managed Agents: what it is, pricing and its limits

What it is

Anthropic’s overview puts it this way: “Instead of building your own agent loop, tool execution, and runtime, you get a fully managed environment where Claude can read files, run commands, browse the web, and run code securely.” The launch post calls it “a suite of composable APIs for building and deploying cloud-hosted agents at scale.” Four things make it up:

• An agent: the model, system prompt, tools, MCP servers and skills. Claude 4.5 and later models are supported, and a session can override the agent’s model.
• An environment: where sessions run, either a cloud sandbox Anthropic manages or, since May 2026, a sandbox on your own infrastructure, where a program you run carries out the bash and file tool calls and Claude still decides them on Anthropic’s side.
• A session: one running instance of the agent doing one task. Each gets “a fresh Linux container,” even when several share an environment.
• Events: what your application and the agent send each other, streamed back as server-sent events and kept server-side.

The built-in tools are bash, read, write, edit, glob, grep, web fetch and web search, all on by default when you include the toolset, plus MCP servers and tools your own application carries out. The cloud sandbox is Ubuntu 24.04 on x86_64 with up to 8 GB of memory and 10 GB of disk, and comes with Python, Node.js, Go, Rust, Java, Ruby, PHP, PostgreSQL and Redis (not started), git, tmux, ffmpeg, LibreOffice and Playwright with Chromium. Packages you list on the environment are installed before the agent starts and cached for later sessions.

Around that core, added since launch: memory stores mounted into the sandbox as folders (up to 8 per session, 10,000 memories each), scheduled deployments that start sessions on a cron schedule to the minute, vaults that attach a credential at the network boundary so the agent “never sees your key,” and a hard spending ceiling per session.

You reach it with an API key from the Claude Console, through Anthropic’s SDKs, the ant command-line tool or the REST API, and every request carries the beta header managed-agents-2026-04-01 (the SDKs add it). The Console has a test runner, and ant beta:sessions connect attaches a terminal to a running session to watch it, send a message or approve a tool call. There is no app for it: it is a building block for products, and Anthropic’s branding rules for partners say a product built on it should not “appear to be Claude Code, Claude Cowork, or any other Anthropic product.”

What it costs

Anthropic’s pricing page: “Claude Managed Agents is billed on two dimensions: tokens and session runtime.”

• Tokens at the ordinary API rates for the model, with the usual prompt-caching prices. Web searches inside a session cost $10 per 1,000. The Batch API’s discount does not apply (“There is no batch mode”).
• Runtime at $0.08 per session-hour, measured to the millisecond and counted “only while the session’s status is running.” Time spent idle, waiting for your next message or for a tool approval, is free, and there is no separate charge for the container.

The token rates, per million tokens, as listed on October 20, 2026:

Anthropic’s own example: a one-hour session on Opus 5.5 that uses 50,000 input tokens and 15,000 output tokens costs $0.58, of which $0.08 is runtime; with 40,000 of the input read from the cache, $0.428.

What a month comes to depends on the work, so here is the arithmetic for one session shape Anthropic publishes, a real coding session of about six minutes from its Claude Code costs page: 1,200 input tokens, 940,000 read from the cache, 50,000 written to it and 5,300 output. Assuming a Managed Agents session doing the same work uses the same tokens and runs for ten minutes:

That is arithmetic from the published rates, not a measurement, and the token lines are the uncertain part: the harness may use more or fewer tokens than Claude Code for the same task, and Anthropic says the tokenizer in Claude 4.7 and later models produces “approximately 30% more tokens for the same text.” The runtime line has a hard ceiling, though: a session running every hour of a 30-day month adds $57.60. So the choice between Managed Agents and a machine of your own is rarely about the $0.08. It is about whether the work is paid for in API tokens at all, since Claude Code on your own machine, signed in with a Pro, Max or Team plan, draws on the plan’s usage limits instead.

The Max and Team credits

Since October 7, 2026, Max and Team plans come with monthly API credits ($100 on Max 5x, $200 on Max 20x, $20 or $100 a Team seat, pooled and capped at $500 a team), and Anthropic’s credits article lists Claude Managed Agents among what they cover, with the Claude API and the Agent SDK. Free, Pro and Enterprise plans are not eligible. You claim them by linking one Claude Console organization to the plan, after seven days on it, and no card is needed. Unused credits expire at the end of each billing cycle, and when they run out with no other credit in the organization, “API requests stop until your next monthly credits arrive. Usage is never charged to your Claude plan.”

On the arithmetic above, Max 5x’s $100 covers a daily job several times over, and a month of working-day coding on Opus goes well past it. The credits never cover interactive Claude Code in the terminal, IDE, desktop or web.

What a session keeps, and for how long

A session survives between your messages: Anthropic’s events page says “when a session goes idle, its sandbox is checkpointed,” with “the filesystem, installed packages, and any files the agent created.” Two limits sit on that:

• Thirty days, from creation. “Sandbox state is preserved for only 30 days after the sandbox is created, and activity does not extend this window.” After that a resumed session starts from a fresh sandbox, with its conversation history still there.
• Nothing shared between sessions. “Sessions do not share filesystem state.” What one session should hand to the next goes through a memory store, the Files API (files written to /mnt/session/outputs/, up to 500 per session) or a GitHub repository mounted into the sandbox.

Deleting a session deletes the files it produced with it. The conversation history is kept until you delete the session.

What you can’t do there

• Reach in. The docs describe outbound network rules only: no port, preview address or shell into a cloud sandbox. A web server, a database another program connects to, or a bot that has to stay up does not fit.
• Sign in with a subscription. It runs on an API key and is billed as API usage or against the Max and Team credits; a plan’s usage limits do not apply to it.
• Use it through Amazon Bedrock or Google Cloud. It is “not available on partner-operated cloud platforms”. It is on Claude Platform on AWS, where a session running with no input for six hours has to be authenticated again.
• Get zero data retention. “Managed Agents is not currently eligible for Zero Data Retention or HIPAA Business Associate Agreement (BAA) coverage.”
• Bring your Claude Code setup as it is. Anthropic’s migration page lists what changes from the Agent SDK: the CLAUDE.md hierarchy becomes one system prompt on the agent, and plan mode, hooks, slash commands and turn limits become your client’s job.

Two defaults to set deliberately. An environment created through the API with no network setting gets unrestricted outbound access, while the Console’s form starts at limited; and per the environments page, “the agent toolset’s default permission policy is always_allow, so bash commands run without approval.”

Managed Agents, Anthropic’s other clouds, and your own machine

Anthropic has three ways to run Claude on its own machines, and the Agent SDK puts the same agent loop on yours. As of October 20, 2026:

Anthropic’s own pointer between the last two shapes runs both ways. Its Agent SDK hosting page says “If you don’t need to run the agent loop itself on your own infrastructure, consider Managed Agents instead,” and the migration page puts the difference in a sentence: “the SDK runs in a process you operate, while Managed Agents runs in Anthropic’s infrastructure.”

When a machine of your own fits better

Managed Agents is the better shape when you are building a product: many isolated sessions on demand, each billed only while it works, with no server to run or secure, scheduling and credential handling built in. A machine of your own is the better shape when the agent is yours rather than your customers’:

• the work should be paid from a Pro, Max or Team plan rather than in API tokens;
• its files, tools and sign-ins should still be there in a month, and one task’s results should be on the same disk as the next;
• something has to keep running or be reachable: a dev server, a database, a bot, a browser signed in to the sites the work touches;
• the job needs more than 8 GB of memory or 10 GB of disk;
• you want Claude Code itself, with your CLAUDE.md, hooks and slash commands, and a session you can open, type into and leave.

That machine has to stay on, which rules out a laptop for scheduled or overnight work; a desktop that never sleeps, a home server or a cloud server does it, and running Claude Code on a schedule on a server covers the setup. Everpod’s developer pod is such a machine: a permanent cloud computer (VPS), set up and secured for you, with your pick of Claude Code, Codex, OpenCode, Pi, Hermes and OpenClaw installed, reached only through your own private network, from $24 a month. A developer pod is one machine that stays yours: what you installed, the servers you left running, Docker and your work in progress are where you left them tomorrow.