n0rds3c.dev / notes / draft
Progressive trust for agentic workflows
. Draft. Not finished; the argument is here, the polish is not.
The question I get most about agentic workflows is some version of “how much should we let it do?” The honest answer is that the number is not fixed. It should move, and it should move based on evidence rather than on how impressive the demo was.
A new colleague does not get production credentials on day one. They get a scoped task, a reviewer, and a way to undo mistakes. If the work holds up, the scope widens. Nothing about that process is specific to people. It is how you extend trust to any actor whose judgment you cannot yet verify, and a model is exactly that.
In practice that means three things. Actions are classified by consequence, not by how the model feels about them. Consequential actions wait for a person until the trail shows a run of clean outcomes at the current scope. And every widening of scope is itself a logged, approved action, so you can see who decided the agent was ready and on what basis.
The failure mode this prevents is the quiet one: an agent that was granted a broad scope on a good day and kept it through every bad one after. Trust that only goes up is not trust. It is a permission you forgot you gave.