Autonomy Is a Dial, Not a Switch
Enterprise agents should earn scope gradually. A framework for staged autonomy: supervised first, shadow mode second, bounded autonomy third.
The false binary
Discussions about agents oscillate between 'chatbot' and 'autonomous employee'. Production reality is a dial: how many actions may the system take without approval, under what budget, with what audit trail? Setting that dial is an engineering decision with measurable inputs.
Stages that work
Stage one: the agent proposes, humans act. Stage two: the agent acts in shadow mode against real workloads, scored offline. Stage three: bounded autonomy on low-risk actions with hard budgets and kill switches. Each stage produces the evaluation data that justifies the next.
Budgets over prompts
Safety enforced by prompt text is theater. Enforce it in the runtime: token budgets, time limits, cost caps, permission-scoped tools, approval gates. The prompt can be jailbroken; the runtime cannot.
Continue