Claude Managed Agents: What They Are and How to Use Them

Claude Managed Agents is Anthropic's fully hosted agent runtime that handles sandboxing, session persistence, tool execution, and tracing so you can run Claude as an autonomous agent without managing infrastructure. Launched in public beta on April 8, 2026, it costs $0.08 per session-hour on top of standard Claude API token pricing.

What Are Claude Managed Agents?

Claude Managed Agents is a suite of APIs on the Claude Platform that handles the infrastructure layer for running AI agents at scale. As of April 2026, it is in public beta.

Instead of building your own execution loop, sandboxing, and tool-calling infrastructure, Managed Agents provides all of that as a service. You define what your agent should do, which tools it can access, and what permissions it has. Anthropic's infrastructure handles secure execution, session state, error recovery, and streaming results back to your application.

The core value proposition is speed to production. Anthropic's own framing is "get to production 10x faster" — and for teams already committed to Claude as their model provider, the claim is credible. The orchestration layer that typically takes weeks to build (retry logic, sandbox isolation, tool routing, session persistence) ships out of the box.

Pricing Breakdown

Claude Managed Agents bills on three axes simultaneously: token consumption, session runtime, and tool-triggered costs. As of April 2026, there is no flat monthly fee.

For a typical 30-minute coding agent session using Claude Sonnet 4, cost analyses estimate total spend between $0.50 and $2.00 depending on token volume. The $0.08/hour runtime component is marginal — most cost comes from tokens.

Architecture: Sessions, Harness, and Sandbox

Managed Agents is built around three core abstractions: the session, the harness, and the sandbox. Understanding these is essential for designing agents that behave predictably.

The session is an append-only log of every event — user messages, Claude responses, tool calls, tool results, and errors. Sessions persist through network disconnections. If your client drops, you reconnect and resume from the last event rather than restarting.

The harness is the execution loop that calls Claude, routes tool calls to the right infrastructure, and handles retries. Anthropic manages this entirely — you do not write orchestration code.

The sandbox is the isolated execution environment where Claude can run code, edit files, and interact with external services. Each session gets its own sandbox with a persistent filesystem. Sandboxes are destroyed when the session ends.

According to Anthropic's engineering blog post on scaling managed agents, the key architectural decision was decoupling "the brain from the hands" — separating Claude's reasoning from the execution infrastructure so both can scale independently.

Getting Started with the API

All Managed Agents endpoints require the managed-agents-2026-04-01 beta header. The official quickstart walks through creating your first agent in under 10 minutes.

The workflow follows three steps: define your agent configuration (system prompt, available tools, permissions), create a session, and stream events. Agent definitions live in the Claude Console or can be created programmatically via the API.

Built-in tools include code execution, file editing, web search, and computer use. You control which tools are available by specifying them in the agent configuration. Custom MCP tools can also be connected for domain-specific workflows.

# Example: Create a session (simplified)
curl -X POST https://api.anthropic.com/v1/managed-agents/sessions \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-beta: managed-agents-2026-04-01" \
  -H "content-type: application/json" \
  -d '{
    "agent_id": "agent_01abc...",
    "messages": [{"role": "user", "content": "Analyze the test failures in this repo"}]
  }'

Managed Agents vs Self-Hosted Agents

Managed Agents and self-hosted frameworks like OpenClaw or Hermes Agent solve the same problem — running autonomous AI agents — but make different tradeoffs around control, cost, and flexibility.

For teams that need multi-model flexibility, complete data isolation, or are already running OpenClaw, self-hosted remains the stronger choice. For teams that want the fastest path to a production Claude agent without infrastructure overhead, Managed Agents is hard to beat. Many production setups use both — Managed Agents for cloud-based tasks, OpenClaw for local or sensitive workloads.

Limitations and Tradeoffs

Managed Agents has real constraints that matter for production use cases. As of April 2026, the public beta does not include multi-agent coordination or self-evaluation — those features require separate research preview access.

The platform is Claude-only. You cannot run GPT-5, Gemini, GLM, or open-source models through it. If your workflow requires model diversity or fallback routing, self-hosted is the only option.

Data flows through Anthropic's infrastructure. Enterprise agreements address many compliance concerns, but for workloads requiring complete air-gapped isolation, Managed Agents is not suitable.

Session sandboxes are ephemeral — they are destroyed when the session ends. If your agent needs persistent storage across sessions, you need to handle that externally.

When NOT to use Managed Agents: if you need multi-model routing, complete data sovereignty, persistent agent state across sessions, or if your cost structure favors fixed infrastructure over pay-per-use at high volume.

Related Guides

• Claude Dispatch vs OpenClaw (2026)
• AI Agent Pricing Compared (2026)
• Best Claude Models for OpenClaw
• Managed OpenClaw Services Compared (2026)

FAQ

How much does Claude Managed Agents cost?

Claude Managed Agents charges $0.08 per session-hour (billed to the millisecond) plus standard Claude API token rates. For Claude Opus 4.6, that means $5 per million input tokens and $25 per million output tokens. Web search costs $10 per 1,000 searches. There is no flat monthly fee — costs scale entirely with usage.

Can I use non-Claude models with Managed Agents?

No. As of April 2026, Claude Managed Agents only supports Anthropic's own Claude models. If you need to run GPT, Gemini, GLM, or open-source models, you need a self-hosted agent framework like OpenClaw or Hermes Agent.

Is Claude Managed Agents the same as the Agent SDK?

No. The Agent SDK is a client-side library for building agentic loops in your own code. Managed Agents is a server-side service where Anthropic runs the agent loop for you. The SDK gives you more control; Managed Agents gives you less setup.

How does Managed Agents handle security?

Each session runs in an isolated sandbox with scoped permissions. You define which tools and resources the agent can access. Secrets are managed through Anthropic's infrastructure and never exposed in session logs. Enterprise accounts get contractual data privacy protections.

When should I use Managed Agents vs OpenClaw?

Use Managed Agents when you want the fastest path to a production Claude agent without infrastructure overhead. Use OpenClaw when you need multi-model support, complete data isolation, persistent agent memory, or when you are already running self-hosted infrastructure. Many teams use both for different workloads.