Managed Agents new

What it is

Most agent frameworks give you a loop and leave you to host it. Managed Agents inverts that: Anthropic hosts the loop, the execution environment and the state, and exposes it as a set of APIs. It launched in public beta on April 8, 2026.

How it works

The composable pieces are:

• Agents: a definition with instructions, tools and a model.
• Sessions: a run, with state that survives interruptions.
• Environments: a managed sandbox where the agent can read files, run commands, browse and execute code.
• Tools: built-in ones plus your own via MCP.
• Permissions: what needs approval before it happens.
• Scheduling: runs on a timer.

Pricing is token usage plus a per-session-hour charge for active runtime. In May 2026 Anthropic added a research preview called dreaming, where the agent reviews past sessions to improve its own memory.

When to use it

You want to ship an agent and not operate its infrastructure. If you need a custom loop, unusual isolation, or a different model provider, run your own with a sandbox provider.

Watch out for

• Beta. APIs are moving.
• Vendor-specific by design. The agent definition does not port.