Claude Managed Agents in Public Beta: Farewell to Sandbox, Orchestration, and DIY State Management

Anthropic has launched Claude Managed Agents in public beta: a suite of APIs designed to build and deploy AI agents in the cloud without having to manage the underlying infrastructure directly. The announcement was shared yesterday evening, April 8, 2026, and the product is already accessible to all API accounts without a separate access request, unlike some features still in research preview such as multiagent, outcomes, and memory.

Claude Managed Agents Solves Agent Management Issues

Bringing an AI agent into production normally requires months of work: secure sandboxes, credential management, state management, tool execution, and the constant need to update everything with every new release of the base model.

Claude Managed Agents shifts the burden to Anthropic, relieving the user: the developer defines the agent's behavior (system prompt, enabled tools, MCP server, skills) and Anthropic takes care of the rest: orchestration, tool execution, context management, and error recovery.

This translates, in practical terms, into a compression of the time-to-production from months to days. Anthropic, in its announcement, cites the experience of some organizations that have already used Claude Managed Agents in preview, such as Rakuten, which implemented specialized agents for product, sales, marketing, finance, and HR in about a week each. Asana, Notion, and Sentry are among other early adopters already operational.

How the Architecture Works

The product revolves around four concepts:

Agent: the model with its system prompt, tools, MCP server, and skills
Environment: a configurable cloud container with pre-installed packages like Python, Node.js, Go, and network access rules
Session: an instance of the agent running on a specific task
Events: the messages exchanged between the application and the agent, including user turns, tool results, and status updates

The operational workflow takes place in five steps. An agent is created once and referenced by ID in all subsequent sessions. The environment is configured with the necessary packages and network permissions. A session is initiated by pairing the agent and environment. From there, textual events are sent, and the agent executes the tools autonomously, streaming results via Server-Sent Events (SSE). The entire event history is persistent on the server and retrievable at any time. It is possible to intervene during execution by sending new events to re-orient the agent or stop it.

Integrated Tools and Optimizations

Agents have native access to a complete set of tools: Bash for executing shell commands in the container, file operations (read, write, edit, glob, grep), web search, fetching URLs, and connecting to external MCP servers. The harness includes built-in optimizations such as prompt caching and context compaction, designed to maintain high output quality on long tasks without inflating costs.

In internal tests, Anthropic measured a 10-point improvement in task success rate compared to the standard prompting approach, with the most significant gains on the most complex tasks. The structure of the harness, specifically optimized for Claude, seems to make a difference compared to generic solutions that operate at the Messages API level.

Two Paths to Build with Claude

Anthropic maintains two distinct paths for developers. The Messages API remains the choice for those seeking granular control over the agent loop, custom state management, and fine integration with existing systems.

Claude Managed Agents is the path for long-running tasks, asynchronous sessions, and those wanting to minimize the infrastructure to manage. These are not competing options: many teams will use both depending on the use case.

To access the Managed Agents, it is sufficient to include the beta header anthropic-beta: managed-agents-2026-04-01 in all requests. The official SDK sets it automatically. The rate limits are set to 60 requests/minute for creation operations (agents, sessions, environments) and 600 requests/minute for read and streaming operations.

Early Adopters and Concrete Use Cases

Notion uses Managed Agents to allow teams to delegate open tasks, from coding to generating slides and spreadsheets, without leaving their workspace.

Sentry has integrated the functionality into its existing debugging tool: the flow goes from identifying the bug to a code fix already ready for review in a single step.

Rakuten has built specialized agents for different business functions, each deployed in about a week, a time unthinkable using the traditional approach.

The CTO of Asana, Amritansh Raghav, stated that the product has enabled his team to release advanced features much more quickly, freeing engineers from infrastructure to focus on user experience. An endorsement that exactly reflects the positioning that Anthropic is pursuing: not a model to integrate, but a platform to build finished products.

Features in Research Preview

Some of the most ambitious features of Managed Agents are still a step behind the public beta and require separate access. Multiagent (the ability of an agent to instantiate and coordinate other agents in parallel), outcomes, and persistent memory are in research preview. Those wanting to test them must explicitly apply to Anthropic. For production use on complex multi-agent architectures, it is advisable to wait for these features to exit preview before building anything critical on them.

Pricing and Availability

Claude Managed Agents is now available in public beta on the Claude Platform, enabled by default for all API accounts. Billing follows the token consumption at the standard Claude rates - $3/MTok in input and $15/MTok in output for Claude Sonnet 4.6, the model currently recommended for most agentic tasks. In addition, the cost of the container runtime is $0.08 per hour for execution in a sandbox environment. For partners integrating Managed Agents into their products, Anthropic has also published branding guidelines: it is allowed to use "Claude Agent" or "{AgentName} Powered by Claude", while references to "Claude Code" or "Claude Cowork" are explicitly prohibited.