Overview
Claude Managed Agents is a fully managed infrastructure service from Anthropic that lets developers build and deploy autonomous AI agents at scale without managing their own agent loops, sandboxes, or tool execution layers. Launched in public beta on April 8, 2026, it provides a pre-built orchestration harness where Claude can read files, run shell commands, browse the web, and execute code inside secure cloud containers.
The service targets development teams and enterprises that need long-running, stateful AI agents for multi-step workflows and asynchronous tasks running on Anthropic-managed infrastructure. Common use cases include coding workflows, document processing, and multi-step automation. Rather than spending weeks wiring up container orchestration, state management, and credential handling, teams can go from prototype to production deployment in days.
Anthropic's internal testing showed Managed Agents improved task success rates by up to 10 points over standard prompting loops, with the largest gains on complex, multi-step problems. Anthropic's launch announcement cites Notion, Rakuten, Asana, and Sentry as early adopters, with Rakuten deploying specialist agents across product, sales, marketing, and finance within a week.
Key Features
Managed Agent Harness — Provides a complete orchestration loop that handles tool routing, context management, and error recovery automatically, eliminating the need to build and maintain custom agent infrastructure.
Secure Sandboxed Execution — Runs agent code inside isolated cloud containers with pre-installed packages (Python, Node.js, Go), configurable network access rules, and mounted file systems, keeping credentials inaccessible to generated code.
Session Persistence & Fault Tolerance — Maintains durable event logs external to any container, enabling crash recovery, mid-session steering, and full event history retrieval without data loss.
Multi-Agent Coordination — Allows agents to spawn sub-agents and delegate subtasks across multiple execution environments, enabling parallel task processing for complex workflows (currently in research preview).
MCP Server Integration — Connects to external tool providers through the Model Context Protocol, letting agents connect to remote MCP servers and external systems such as GitHub, Slack, databases, and custom internal tools.
Built-in Performance Optimization — Includes built-in prompt caching, compaction, and other performance optimizations for higher-quality, more efficient agent outputs.
How to Get Started
Getting started with Claude Managed Agents requires a Claude API key and the beta header managed-agents-2026-04-01 on all requests (the SDK sets this automatically).
The setup follows four steps:
Create an Agent — Define the model, system prompt, tools, MCP servers, and skills. Create the agent once and reference it by ID across sessions.
Create an Environment — Configure a cloud container template with pre-installed packages, network access rules, and any files you want mounted.
Start a Session — Launch a session referencing your agent and environment configuration. The harness begins running immediately.
Send Events and Stream Responses — Send user messages as events. Claude autonomously executes tools and streams back results via server-sent events (SSE). You can steer or interrupt the agent mid-execution by sending additional user events.
Certain advanced features including outcomes, multi-agent coordination, and memory are available in research preview and require a separate access request through the Claude Platform.
Pricing & Plans
Claude Managed Agents uses a usage-based pricing model billed on two dimensions: token consumption and session runtime.
Session Runtime:
| Component | Rate | Details |
|---|---|---|
| Session runtime | $0.08 per session-hour | Billed only while session status is running; idle, rescheduling, and terminated time is free |
Token Costs (standard Claude API rates):
| Model | Input | Output |
|---|---|---|
| Claude Opus 4.6 | $5 / MTok | $25 / MTok |
| Claude Sonnet 4.6 | $3 / MTok | $15 / MTok |
| Claude Haiku 4.5 | $1 / MTok | $5 / MTok |
Additional Costs:
- Web search inside sessions: $10 per 1,000 searches
- Web fetch: No additional cost beyond token usage
- Prompt caching: Cache reads cost 10% of standard input price
Cost Example: A one-hour coding session using Claude Opus 4.6 consuming 50,000 input tokens and 15,000 output tokens costs approximately $0.71 total ($0.25 input + $0.375 output + $0.08 runtime). With prompt caching active, the same session drops to around $0.53.
Rate Limits:
- Create endpoints: 60 requests per minute per organization
- Read endpoints: 600 requests per minute per organization
- Standard Claude API tier-based rate limits also apply
Note: Batch API, fast mode, and data residency modifiers from the Messages API do not apply to Managed Agents sessions. Enterprise customers can contact Anthropic's sales team for custom pricing and higher rate limits.
How It Compares
| Feature | Claude Managed Agents | Other managed agent platforms | Self-hosted frameworks (n8n/LangChain) |
|---|---|---|---|
| Setup time | Days | Weeks | Weeks to months |
| Infrastructure | Fully managed | Fully managed | You manage everything |
| Model support | Claude models only | OpenAI models only | Any model |
| Session persistence | Built-in | Varies | Manual implementation |
| Multi-agent support | Research preview | Limited | Custom implementation |
| Pricing model | Token + runtime hourly | Value-based | Free software + compute costs |
| MCP integration | Native | Not available | Plugin-dependent |
Claude Managed Agents differentiates primarily on managed infrastructure, secure sandboxing, state persistence, and Anthropic-native orchestration. Its architecture decouples the reasoning model from execution containers, delivering significantly faster time-to-first-token than tightly coupled alternatives. The trade-off is vendor lock-in: Claude Managed Agents runs only Claude models through Anthropic's direct Claude API, so moving to another model provider would require reworking your integration and runtime assumptions.
Self-hosted solutions like n8n or LangChain offer model flexibility and no runtime fees, but require teams to manage their own container orchestration, state persistence, and security infrastructure. For teams already committed to Claude's model family, Managed Agents removes substantial infrastructure overhead.
Best For
- Development teams building AI-powered automation pipelines who want to skip months of infrastructure setup
- Enterprises deploying customer support, data processing, or document analysis agents at production scale
- Companies already using Claude's API — including Claude Code users — that need long-running, stateful agent sessions with crash recovery
- Startups and agencies building multi-step AI workflows who prioritize speed to market over model flexibility
- Engineering teams automating code review, bug triage, and pull request workflows with autonomous agents
FAQ
What is Claude Managed Agents?
Claude Managed Agents is a managed infrastructure service from Anthropic that provides a pre-built agent harness, secure sandboxed execution environments, and session management for running autonomous AI agents in the cloud. It handles tool execution, state persistence, and error recovery so developers can focus on defining agent behavior rather than building infrastructure.
How much does Claude Managed Agents cost?
Pricing has two components: session runtime at $0.08 per hour (billed only while actively running) and standard Claude API token rates. A typical one-hour session using Opus 4.6 with moderate token usage costs under $1. Web search inside sessions costs $10 per 1,000 searches. There is no monthly subscription fee.
Is Claude Managed Agents available now?
Yes, Claude Managed Agents launched in public beta on April 8, 2026 and is enabled by default for all Claude API accounts. You need to include the managed-agents-2026-04-01 beta header on API requests. Advanced features like multi-agent coordination and memory require a separate access request.
Can I use models other than Claude?
No. Claude Managed Agents exclusively runs Anthropic's Claude models (Opus 4.6, Sonnet 4.6, Haiku 4.5, and others). It does not support non-Claude models and is available only through Anthropic's direct Claude API.
What tools can agents use?
Agents have access to bash commands, file operations (read, write, edit, glob, grep), web search and fetch, and MCP server connections for external integrations. You can connect agents to remote MCP servers and use Anthropic-managed tools.
How does session persistence work?
Sessions maintain a durable, append-only event log stored externally from execution containers. Event history is persisted server-side, and the managed harness is designed for long-running sessions with durable progress tracking. Runtime is only billed while the session is actively running.
Is my data secure?
Agent code runs in isolated sandboxed containers. Credentials are stored in secure vaults accessed via MCP proxy and are never exposed to Claude's generated code. Git tokens are bundled during sandbox initialization with scoped permissions.
What is multi-agent coordination?
Multi-agent coordination allows a primary agent to spawn sub-agents that work on subtasks in parallel across separate execution environments. This feature is currently in research preview and requires an access request through the Claude Platform.
