85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Glossary

What is a meta-harness? Brain vs hands explained

A meta-harness is an agent platform built from stable interfaces that separate the brain (model and harness), the hands (sandboxes and tools) and the session log, so different harnesses can run on it and each part can fail independently.

Anthropic described the design in "Scaling Managed Agents: decoupling the brain from the hands" (2026-04-08). The brain is Claude plus its harness, the loop that calls the model and routes tool calls. The hands are the sandboxes and tools that carry out actions. The session is an append-only log of everything that happened. Each is an interface that assumes little about the others.

The post calls Managed Agents a meta-harness: opinionated about the interfaces around the model, but not about which harness runs inside them. Claude Code could run on it, and so could a harness written for one narrow task.

Why the parts were split

The first version ran the session, harness and sandbox in one container. That had real benefits, such as file edits as direct system calls and no service boundaries, but the container became a pet rather than cattle. If it failed, the session was lost. Debugging needed shell access to containers that often held user data. Connecting to a customer's private network meant peering or running the harness inside the customer's environment.

The security problem was sharper. Generated code ran next to credentials, so a successful prompt injection could expose tokens.

The interfaces

  • One call for every hand

    Any tool, MCP server or container is reached through the same shape: execute a named action with an input and get a string back. A failed container becomes a tool error the brain can retry.

  • Provision on demand

    Containers are created from a standard recipe only when a session needs one, rather than before the first token.

  • The session outside the window

    The log lives outside both harness and sandbox. A crashed harness wakes from it, and the brain can read slices of it or reread events before acting, so nothing is lost to compaction.

  • Credentials out of reach

    A repository token is used once at sandbox start to clone and configure the remote. MCP OAuth tokens stay in a vault, and tool calls go through a proxy that attaches them, so the harness never sees a credential.

What changed, and the costs

Anthropic reports that after decoupling, median time to first token fell by roughly 60% and p95 by more than 90%, because sessions no longer wait for a container they may not need.

The post is open about the costs. Working across several hands is harder for the model than working in one shell, and splitting context storage from context management leaves the harness responsible for deciding what each call sees.

How Motiontic's engineers apply it

The pattern does not require a hosted platform. Motiontic's engineers apply the same three seams on self-hosted agents: tool execution behind one narrow interface, a session log in a database rather than in process memory, and secrets held by a proxy the agent's code cannot read.

Those seams also make the system testable. A sandbox can be replaced with a fake in evals, and a run can be replayed from its log when a failure needs explaining.

A coupled agent container compared with a meta-harness

A coupled agent container compared with a meta-harness
Coupled containerMeta-harness
Container failureSession lostRetryable tool error
Where history livesInside the containerIn a durable session log
CredentialsNext to generated codeIn a vault behind a proxy
Start-upWait for the containerContainer provisioned only when needed
Harness choiceOne harness, built inAny harness that fits the interfaces

Common questions

What does brain vs hands mean for AI agents?

The brain is the model and the harness that runs its loop. The hands are the sandboxes and tools that act. Keeping them separate means a broken sandbox does not end the run and credentials stay away from generated code.

Is a meta-harness the same as an agent harness?

No. An agent harness runs one agent's loop, tools and context. A meta-harness is the layer of interfaces underneath that many different harnesses can run on.

Who uses the term meta-harness?

Anthropic used it in its April 2026 engineering post on Managed Agents to describe a system that is opinionated about the interfaces around Claude but not about the specific harness.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.