AI agents deserve a work budget and a permissions profile
The practical choice: give an agent a limited work run
The conclusion is straightforward: let an agent execute only when the task, permissions, stop rule, and human owner have been defined in advance within a single work run. For other tasks, the agent should continue to prepare, flag issues, or stage a change. This distinction prevents a convincing answer from being mistaken for permission to act. The supplied permissions passage describes the risk of sensitive actions being performed without the user's intent or authorization and notes that users have different needs and preferences. Its scope is agent-permission design; the passage does not prove that every organization needs the same permissions structure. The passages do not support a general guarantee of safety, cost, or quality: the cost findings come from agentic coding tasks on SWE-bench Verified, and the other source is a survey of proposals for user permissions.
Make cost a stop rule, not an afterthought
For agentic coding tasks, budget is an operational control point. The supplied study analyzes eight frontier models on SWE-bench Verified and reports that such tasks can consume many tokens, that runs on the same task can vary widely, and that more tokens do not automatically lead to higher accuracy. This is not a general price forecast for all agents or vendors, but it is sufficient reason to set a cap and a stopping point for each run. For every defined task, therefore, measure the objective, the input tokens used, the outcome, and the point at which the agent must stop or involve a human. For each work run, record: task objective, permitted action, owner, required permissions, human approval, token or cost cap, stop rule, and recovery step.
A decision card for the first production run
Use this decision card for every candidate task: describe the permitted action; note which user or role grants permission; choose a human check before an irreversible step; set a token or cost cap; and specify what happens in the event of an overrun or uncertainty. Start with a review or preparation step, and adjust execution permissions only after the run has proven auditable. This is not evidence that a particular technical setup is safe or inexpensive; the passages address, respectively, agentic coding tasks on SWE-bench Verified and a survey of 21 proposals for permission systems. They do not provide individualized legal, financial, security, or professional advice.
