Claude Managed Agents Explained: What It Is, How It Works, Pricing, and When to Use It (2026)

What is Claude Managed Agents?

Anthropic's documentation defines it as a "pre-built, configurable agent harness that runs in managed infrastructure. Best for long-running tasks and asynchronous work." In plainer terms: the agent loop that powers Claude Code, plus a sandbox for it to work in, plus durable session storage, offered as an API. You send a task; Anthropic runs the agent until it is done or needs you.

It launched on April 8, 2026 and is still in public beta: every endpoint requires the managed-agents-2026-04-01 beta header, though access is on by default for all API accounts. If the vocabulary is new, two earlier posts help: an agent harness is everything around the model that lets it act, and an agent runtime is the production layer that keeps it running. Managed Agents is Anthropic selling you both, pre-assembled.

How does Claude Managed Agents work?

There are four objects:

Anthropic's engineering post, Scaling Managed Agents: Decoupling the brain from the hands, explains the design: the "brain" (Claude plus the harness) is separated from the "hands" (sandboxes and tools) and from the session log, so each can fail, scale, and be replaced independently. Anthropic reports that after this change "p50 TTFT dropped roughly 60% and p95 dropped over 90%."

On top of that sit the features you would otherwise build yourself: a built-in toolset (bash, read, write, edit, glob, grep, web fetch, web search), custom tools your code executes, MCP servers with OAuth credentials kept in vaults, skills, multi-agent coordination, scheduled (cron) deployments, per-session budgets, permission policies, outcome rubrics, and the ability to steer or interrupt a running agent. The harness also handles prompt caching and context compaction for you.

Claude Managed Agents quickstart

The official quickstart creates the three objects and streams the result:

# From Anthropic's Managed Agents quickstart (Python), trimmed.
# Every Managed Agents endpoint needs the managed-agents-2026-04-01 beta header.
agent = client.beta.agents.create(
    name="Coding Assistant",
    model="claude-opus-5-5",
    system="You are a helpful coding assistant. Write clean, well-documented code.",
    tools=[{"type": "agent_toolset_20260401"}],
)

environment = client.beta.environments.create(
    name="quickstart-env",
    config={"type": "cloud",
            "networking": {"type": "limited", "allow_package_managers": True}},
)

session = client.beta.sessions.create(
    agent=agent.id, environment_id=environment.id, title="Quickstart session",
)

with client.beta.sessions.events.stream(session.id) as stream:
    client.beta.sessions.events.send(session.id, events=[{
        "type": "user.message",
        "content": [{"type": "text", "text": "Create a Python script that generates the first 20 "
                     "Fibonacci numbers and saves them to fibonacci.txt"}],
    }])
    for event in stream:
        ...  # agent.message / agent.tool_use / session.status_idle

There are SDKs for Python, TypeScript, Java, Go, C#, Ruby, and PHP, a CLI (ant apply agent.md lets you define an agent in a Markdown or YAML file), and in Claude Code the command /claude-api managed-agents-onboard walks you through setup.

How much does Claude Managed Agents cost?

Per Anthropic's pricing page, you pay for two things, "tokens and session runtime":

Anthropic's worked example: a one-hour session on Opus 5.5 that uses 50,000 input and 15,000 output tokens costs $0.20 + $0.30 + $0.08 = $0.58, or about $0.43 with caching. The runtime fee is small; tokens dominate. That makes model choice and context management the real cost levers, the same ones we cover in LLM routing. Set a per-session budget before you let an agent run unattended.

Claude Managed Agents vs Agent SDK vs AgentCore vs Google's agent platform

The sharpest line is who runs the loop. Anthropic's Agent SDK overview puts it plainly: the SDK runs the agent "in a process you operate," while Managed Agents means Anthropic hosts the agent. The second line is model choice. Bedrock AgentCore advertises "any framework, any model" (and lists the Claude Agent SDK as a supported framework), and Google's agent platform hosts ADK, LangChain, LangGraph, and LlamaIndex agents. Managed Agents is Claude-only. If you build with an open-source framework, see our agent frameworks comparison.

Limits and what to watch

Anthropic's launch customers give a sense of the target: Notion, Rakuten, Asana, Sentry, and Atlassian among them, with Sentry running a debugging agent and a patch-writing agent. That is the sweet spot: well-defined, asynchronous jobs that take minutes to hours, where the team wants results, not agent infrastructure. Pair it with sensible guardrails on tool calls.

When should you use Claude Managed Agents?

Frequently Asked Questions

References

• • Anthropic — Claude Managed Agents overview (docs)
• • Anthropic — Managed Agents quickstart
• • Anthropic — Managed Agents tools
• • Anthropic — Managed Agents API reference and limits
• • Anthropic — Pricing: Claude Managed Agents
• • Anthropic — Claude Managed Agents launch (Apr 8, 2026)
• • Anthropic Engineering — Scaling Managed Agents: Decoupling the brain from the hands
• • Anthropic — Agent SDK overview (SDK vs Claude Code vs Managed Agents)
• • AWS — Amazon Bedrock AgentCore
• • Google Cloud — Agent Engine overview
• • VentureBeat — Claude Managed Agents and lock-in risk

Pricing and beta status verified against Anthropic's docs on October 20, 2026.