Harness
The deterministic scaffolding around the work on a change, whoever does it: compiler errors, formatters, linters, checks, permission gates, typed contracts, leases, graph-backed answers and hooks, placed at points a person chose, with judgment covering the steps between them. It holds a person to the green path the same way it holds an agent (A harness is not an agent thing). Around a model's agent loop, magus supplies one (guard hooks, skills, MCP tools, job leases) and keeps it outside the host's loop. Some vendors use the word for the loop itself; magus keeps the two apart.
- Is: code and reviewed text that answer or decide the same way every time.
- Is not: the model, the agent loop that reads a tool call and runs it, an agent framework, or a prompt.
- Layers, from the inside out: the model; the host's agent loop (Claude Code, Codex, Cursor, OpenCode); the harness magus supplies; magus's deterministic workspace model, which the harness reads.
Elsewhere
Anthropic calls Claude Code and the Claude Agent SDK an agent harness: the loop that sends the model a prompt, runs the tools it calls and feeds back the results. A test harness is the scaffolding that drives code under test. magus means neither: its harness is the deterministic checks around a host's loop, and the loop stays the host's.
Example
magus describe harness shows what a wired host needs: the guard hooks, the skills and the
MCP tools magus supplies.
magus describe harness
Used in
In the docs:
- The agent harness benchmark
- Harness-effectiveness benchmark
- Harness-effectiveness benchmark
- 6. Harness and host configuration
- The outer harness stays outside
- A harness is not an agent thing