On harnesses, contracts, and the shape of autonomy
The model does not decide the terms of its autonomy.
What it sees, which tools it calls, what it spends, and when a person takes over belong to the harness. The same weights serve different systems.
A harness is a contract made executable. It defines permissions, budgets, rules that must hold, and conditions for returning control. Written down, the contract can be tested. Left implicit, it is renegotiated on every run.
Contracts make failure legible. Did the agent exhaust its budget? Fail verification? Need authority it did not hold? If the system cannot tell you, someone must reconstruct the answer from a transcript.
Wider permissions create more room to act and more ways to fail. Tighter constraints limit what proceeds unattended. The engineering lies in choosing those boundaries and making them hold.