# Brain-Hands Decoupling

> Separating the model-and-harness layer that decides what to do from the sandbox, tools, or systems that perform actions.

- Category: Agents & Autonomy
- Stage: Emerging
- Also called: Harness-compute separation, Brain/hands split, Model-sandbox separation
- Canonical: https://www.joelcsaji.com/ai-dictionary/brain-hands-decoupling

## Plain English

The agent's decision loop does not have to live inside the same container or environment as the code it runs.

## Example

A hosted coding agent keeps a durable session log outside the sandbox, then provisions a fresh execution environment only when a tool call needs it.

## Why it matters

The split improves durability, scaling, debugging, and credential isolation for long-running agents.

## Related

- [Agent Harness](https://www.joelcsaji.com/ai-dictionary/agent-harness)
- [Sandbox](https://www.joelcsaji.com/ai-dictionary/sandbox)
- [Durable Execution](https://www.joelcsaji.com/ai-dictionary/durable-execution)

## Sources

- [Anthropic: Managed Agents](https://www.anthropic.com/engineering/managed-agents)
- [OpenAI: Agents SDK Evolution](https://openai.com/index/the-next-evolution-of-the-agents-sdk/)