Anthropic separates an AI agent's brain from its hands

On April 8, 2026, Anthropic launched Claude Managed Agents. It is a hosted service on the Claude platform for AI agents that work over longer periods. An AI agent is a program that can carry out a multistep task on its own. According to Forklog, the agents can use tools, run code, edit files and talk to external services. The service can also resume after failures.

At the same time, Anthropic published an engineering post on how the service is built.

Three parts instead of one lump

The architecture consists of three parts. The "brain" is Claude plus the loop that decides the next step and passes tool calls along. The loop is called a harness. The "hands" are the sandboxes and tools that carry out the actions. A sandbox is an isolated environment where code can run. The "session" is a durable log of everything that happens, and it is append-only.

The session lives outside the harness. This makes it possible to recover the course of events and fetch context again. The hands are treated as interchangeable resources, and the brain should not assume where they run.

From pets to interchangeable parts

Anthropic describes the first model as tightly coupled, treating the infrastructure as "pets," meaning named, hand-tended units. When a container failed, the session was lost. Debugging required access to the container's command line, which exposed user data and created security risks. The model also assumed that resources were located close to the container, which limited flexibility, for example for customers' own VPCs (private networks in the cloud).

In the new design, credentials are never available in the sandboxes where Claude's code runs. Git tokens are used to fetch code but are kept outside the sandbox. OAuth tokens sit in a vault, and calls to MCP tools go through a proxy that fetches them, so they are hidden from the harness.

Faster responses, says Anthropic

Anthropic says time to first token dropped by roughly 60 percent at the median and by more than 90 percent for the slowest 5 percent (p95). This is because the model can start working without waiting for a container to be made ready. The figures are Anthropic's own, and it is not stated how they were measured.

Assumptions go stale

Anthropic writes that a harness contains assumptions about what Claude cannot do on its own, and that they should be challenged regularly because they go stale as the models improve. According to Anthropic, the brain, hands and session evolve at different paces and should therefore be replaceable independently.

Forklog writes that Anthropic is positioning agents as an autonomous layer similar to an operating system, one that can withstand model changes. The outlet also points out that the session log can work around limits on context length. Pricing and customers are not mentioned in the sources.