Managed Agents

0:02 Claude Managed Agents is a hosted harness for running Claude as an autonomous agent: Anthropic runs the agent loop, executes the tools and provides a fresh Linux sandbox for every session, so you write no loop or tool runner of your own. It costs normal model token prices plus $0.08 per session-hour, counted only while a session is actually running (as of Nov 6, 2026). It launched in beta on Apr 8, 2026 and is still beta, with the managed-agents-2026-04-01 header on every call (the SDKs set it).

2:19 Anthropic presents it as one of two ways to build on the Claude API: the Messages API for direct model access and your own loop, and Managed Agents for long-running, asynchronous work on Anthropic's infrastructure (Intro to Claude).

Quick facts

Agents, environments, sessions and events

5:44 Managed Agents has four core ideas (Quickstart):

• Agent: a reusable, versioned configuration (model, system prompt, tools, MCP servers, skills), created once and referenced by ID. A session uses its latest version unless you pin one.
• Environment: the sandbox template. Each session gets its own isolated Linux container; packages preinstalls software with apt, cargo, gem, go, npm or pip, and networking is limited (only listed hosts) or unrestricted (Environments). An API request that leaves out networking gets unrestricted, while the Console form starts at Limited.
• Session: one run of an agent in an environment.
• Events: what flows between your app and the session, such as user.message in, and agent.message, agent.tool_use and session.status_idle streamed back over server-sent events.

The quickstart's Python, trimmed to its four calls:

from anthropic import Anthropic

client = Anthropic()

agent = client.beta.agents.create(
    name="Coding Assistant",
    model="claude-opus-5-5",
    system="You are a helpful coding assistant.",
    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)

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": "Write fibonacci.py and run it"}],
    }])
    for event in stream:
        if event.type == "session.status_idle":
            break

Agents can also be defined as files and created with the ant CLI's ant apply, which records IDs in claude-lock.json, and ant beta:sessions connect attaches a terminal to a live session (Release notes). The Console has a visual agent builder that shows the equivalent API request and an inline test runner (Build in Console); see Claude Console.

Vaults, outcomes, schedules and more

• Vaults hold each end user's credentials (MCP OAuth tokens, bearer tokens, environment variables kept as placeholders in the sandbox), registered once and referenced by ID when a session starts (Vaults).
• Outcomes describe what the finished work should look like. A separate "grader" with its own context checks the result against your Markdown rubric, and the agent keeps going until it passes or hits its iteration limit (Define outcomes).
• Scheduled deployments start sessions on a cron expression in a time zone, at most once a minute (Scheduled deployments).
• Multiagent orchestration lets an agent hand work to subagents, each in its own thread, or to dynamic workflows (beta since Oct 9, 2026) that run many agents in the background (Multiagent orchestration).
• Session budgets (since Aug 7, 2026) set a hard ceiling in US cents: "125" means $1.25, and the session pauses with budget_reached when it gets there (Session budgets).
• Memory stores (beta) keep text documents across sessions, mounted as a directory in the sandbox; dreams, a research preview, tidy a memory store using past transcripts (Memory).
• Self-hosted sandboxes (since May 19, 2026) run the tool calls on your own machines while Claude itself runs at Anthropic.
• Permission policies decide whether each tool call runs at once, waits for approval, or (since Sep 10, 2026) is judged by the server with auto.

How much does it cost?

You pay the model's normal token prices, with prompt caching always on, plus the runtime charge. Anthropic's own worked example: a one-hour Claude Opus 5.5 session with 50,000 input and 15,000 output tokens costs $0.20 + $0.30 + $0.08 = $0.58, or $0.428 when 40,000 of the input tokens are cache reads. Idle, rescheduling and terminated time is free; web search inside a session costs the usual $10 per 1,000 searches; there is no batch discount. The runtime fee replaces the container-hour billing of the Code Execution Tool. Max, Ultra and Team subscribers' monthly API credits cover Managed Agents. See Claude API Pricing.

When to use Managed Agents, and when not

1:41 Anthropic positions it for tasks that run minutes or hours, need a sandbox, keep state across a session, or run on a schedule (Overview). The other options:

25:46 Its limits are real: everything is beta, sessions are not covered by zero data retention because history, sandbox state and outputs are stored at Anthropic, and the agent only runs on the three Anthropic-operated platforms listed above. Teams already committed to a cloud provider's agent service, or needing data to stay in their own systems, may prefer the Agent SDK on their own infrastructure, or self-hosted sandboxes for the tools.

Frequently asked questions

How much does Claude Managed Agents cost? Claude Managed Agents costs the model's token prices plus $0.08 per session-hour while the session is running, metered to the millisecond; idle time is not charged (as of Nov 6, 2026).

Is Claude Managed Agents the same as the Claude Agent SDK? No. The Claude Agent SDK runs the agent loop in your own process and on your own machines, while Claude Managed Agents runs the loop, tools and sandbox on Anthropic's infrastructure.

Is Claude Managed Agents available on Amazon Bedrock? No. As of Nov 6, 2026, Claude Managed Agents runs only on the Claude API, Claude Platform on AWS and Claude Models on Google Cloud (Anthropic managed).

Can a Claude managed agent run on a schedule? Yes. Scheduled deployments start a Claude Managed Agents session on a cron expression, with each run beginning from a user message or a defined outcome.

Does Claude Managed Agents support MCP servers? Yes. A Claude managed agent can list remote MCP servers in its definition, with per-user credentials kept in vaults; see Model Context Protocol.

Sources

• What is Claude Managed Agents?, Claude, Apr 9, 2026 (video)
• Ship your first Managed Agent, Claude, May 26, 2026 (video)
• Claude Managed Agents overview, Anthropic docs, read Nov 6, 2026
• Get started with Claude Managed Agents, Anthropic docs, read Nov 6, 2026
• Build in Console, Anthropic docs, read Nov 6, 2026
• Environments, Anthropic docs, read Nov 6, 2026
• Vaults, Anthropic docs, read Nov 6, 2026
• Define outcomes, Anthropic docs, read Nov 6, 2026
• Scheduled deployments, Anthropic docs, read Nov 6, 2026
• Multiagent orchestration, Anthropic docs, read Nov 6, 2026
• Session budgets, Anthropic docs, read Nov 6, 2026
• Memory, Anthropic docs, read Nov 6, 2026
• Pricing, Anthropic docs, read Nov 6, 2026
• Rate limits, Anthropic docs, read Nov 6, 2026
• Features overview, Anthropic docs, read Nov 6, 2026
• Intro to Claude, Anthropic docs, read Nov 6, 2026
• API credits for Max, Ultra, and Team plans, Anthropic docs, read Nov 6, 2026
• Release notes, Anthropic docs, read Nov 6, 2026