## Configuration

BaseExecution engine such as `codex`, `claude-code`, or `hermes`.

Purpose: The product job this runtime is allowed to perform.

Instructions: Task rules, output contract, role mismatch behavior, and safety boundaries.

Tools and skills: Optional MCP servers, custom skills, and built-in document capabilities.

Model policy: The allowed models and default model fetched from the API.

## Reuse or create

- Reuse an existing configured agent only when purpose, data access, tools, permissions, and output contract match.
- Update it when it is the canonical configured agent for the same feature.
- Create a new configured agent when the risk boundary or output contract changes.

**Open standard**

The base, instructions, tools, skills, and model policy you combine here populate the open harness object — fields like systemPrompt, mcpServers, and disabledTools — that the Unified Harness Protocol specifies as a first-class, addressable configuration. See [the Harnesses chapter of UHP](https://unifiedharnessprotocol.org/spec/harnesses).

[Create API Key →](https://app.harnessrouter.ai/quickstart?ref=docs-create-api-key)

Ready to run this against a live workspace? Keys take under a minute.