Leash and harness are often used interchangeably to describe controls around a system. Confusing them leaves the operator doing work the agent was meant to take on.
A leash limits and directs movement, keeping the operator close enough to correct the next step. A harness protects while allowing movement. It may catch or redirect, but it does not prescribe a step-by-step path.
That is the difference between a bot sitter and a safely bounded agent.
A leash makes the operator part of the loop
A bot sitter is not doing anything wrong. Sometimes a human in the loop is the responsible design. It does not create much autonomy. If a person must steer every read, plan, edit, and retry, the bot is just a faster way to issue instructions while the operator still carries the cognitive load.
A leash makes supervision the mechanism of safety.
That is appropriate for work that requires direct guidance: approving a destructive action, sending a customer message, or applying a production migration. It should not be the permanent model for routine work.
A harness makes useful work safe
A harness answers a different question: what does the system need to function safely?
For an agent, the answer is ordinary engineering. Authority is scoped. Tools have narrow interfaces. Actions have recovery paths. The run has limits, verification, and a record of what happened.
Those controls do not tell the agent which file to inspect or command to run. They define where it can work, what it must prove, and when it must stop. A failed test becomes evidence for the next bounded attempt, not an immediate request for a person to take the wheel.
The physical distinction is useful
When I worked in water filtration, we literally leashed cords. The goal was to route and contain the mess. In confined spaces or at precarious heights, we used a harness. The harness did not tell us how to do the job. It kept a slip from becoming a fall.
Agent systems need both, in the right places. Leash the irreversible decision. Harness the ordinary work around it. Let the system inspect a repository and prepare a patch, then run a bounded test suite and recover from expected failures.
The design test
Do not ask whether an agent is controlled. Ask where the control lives.
If safety depends on a person directing the next action, you have a leash. If it comes from explicit boundaries that let the system work, fail, recover, and show its evidence without exceeding its authority, you have a harness.
The goal is not less control. It is control that scales with the work.
Previous: Agents Are Permission Systems
Next: Inside the AI Harness
The one-line version
A leash keeps an operator in every move; a harness makes autonomous work safe enough to be useful.