What is a supervision harness?
Carlos Lebre Ribeiro ·
A supervision harness is the layer above an AI agent that decides, independently of the agent, which of its actions may run. It checks each action before it happens, clears the safe ones, stops the rest or sends them to a person, and keeps a record of every decision.
How it differs from an agentic harness
An agentic harness is the runtime an agent works in. It separates instruction from enforcement: the prompt says what to do, and the harness makes sure the workflow cannot move on until the required step has happened.
A supervision harness sits one level up. It is independent of the executor, it spans models and frameworks, and it produces the audit trail as its output rather than as a by-product.
Why it has to be independent
An executor grading its own actions is homework marked by the student. No auditor audits their own work, and the same rule applies to an agent. A hyperscaler can build a good classifier, and it still cannot be the independent supervisor of its own agents.
The two questions
Every action has to pass two questions before it runs:
- Is it safe? The agent is competent and uncompromised.
- Is it right? The action complies with your policy and the regulation that applies to you.
The questions come from David Marquet, who used them to take a US Navy submarine from worst to first. Control was handed to the crew only once each action had passed both.
What it governs
A supervision harness governs the business action: the payment, the claim decision, the disclosure. It matters where a high volume of consequential decisions gets made. If the volume is low, a person can review each one. If the stakes are low, a basic check will do.