What is an agentic harness?
An agentic harness is everything around a language model that makes an AI agent reliable and safe in production: context and tools, guardrails and permissions, evaluation and monitoring, and governance. The model is the engine, the harness keeps it on course.
- GuardrailsInput and output checks on every run
- PermissionsMinimal access per system
- EvaluationsA golden set of real cases, with every change
- ObservabilityEvery run traceable, including cost and speed
- Human in the loopApproval moments per risk level
- Version controlEvery change tested, rollback always possible
- Token budgetA cost ceiling per agent and per run
- Emergency stopOne button that shuts the agent down
Explanation
On its own, a language model only produces text. The harness turns it into an agent that uses tools, remembers where it is and stays within limits. Coding agents such as Claude Code and Codex are themselves harnesses around a model.
For every agent that goes live, we check eight components:
- Guardrails: input and output checks on every run
- Permissions: minimal access per system
- Evaluations: a golden set of real cases, with every change
- Observability: every run traceable, including cost and speed
- Human in the loop: approval moments per risk level
- Version control: every change tested, rollback always possible
- Token budget: a cost ceiling per agent and per run
- Emergency stop: one button that shuts the agent down
A good harness is model-independent. When a better or cheaper model comes out, you swap the engine without rebuilding the rest.
What it looks like in practice
At a bank, an agent that prepares client files runs inside a harness with read-only access, a golden set of hundreds of files, logging of every step and a cost ceiling per file. A compliance officer can stop it at any moment.
