AI agents now get a budget and a rulebook for acting on their own — do the dollar limits and the rulebook actually line up?

This explores how the 'autonomy budget' from agent interface design lines up with the scope limits OpenAI's Model Spec writes down for agents. They fit together closely, because they are two ends of one idea. The Model Spec handles the policy side. It proposes a ScopeOfAutonomy record that lists the tools an agent may use, how much it may spend, which high-risk actions are off-limits, and when its permission ends How can autonomous agents act safely within defined boundaries?. It borrows the security rule of least privilege: give the agent only the access the task needs, so you don't have to ask the user to confirm every step. The design community handles the user's side. The AI UX Playground pattern library treats autonomy budgets as one control among several, next to human-in-the-loop checkpoints, suggest/confirm/execute modes, and per-capability permissions Can autonomy controls stay distinct rather than collapse into one toggle?.

Roughly, the pieces map like this. The spec's cost limits are the autonomy budget. Its list of allowed tools matches per-capability permissions. Its banned high-risk activities match the checkpoints where a human has to approve. Its ending conditions have no clear match in the design patterns, and that gap matters. A budget says how much an agent may do. A scope record also says when its permission runs out. Without an expiry, a permission granted for one task can quietly carry over to the next. One difference is worth noticing: the same design community also describes autonomy as a single continuous slider (suggest, draft, execute) Should agent autonomy be a continuous slider or fixed levels?. The Model Spec's record has several separate fields, which points away from a single dial and toward separate controls for each kind of limit.

The surprise is that strict limits can backfire. In one study, models were given a $10 budget and offered a $10.01 purchase. 92% of humans went ahead, but every AI model refused. The models only behaved more sensibly after seeing the reasoning behind the human choices, not just the choices themselves Do AI models follow budgets more rigidly than humans do?. So writing a cost limit into a scope record doesn't settle what the agent will do. Models may treat the number as a hard wall, where a person would read it as a rough guide. Anyone designing autonomy budgets has to decide whether a limit is strict or flexible, and say so explicitly. The Model Spec record, as described here, doesn't say how to handle that.

A second question is who sets the scope, and when it grows. Yocco argues autonomy should expand in three phases (analysis only, then low-risk actions, then full autonomy), and each step should depend on evidence that users trust the agent, not on a schedule Should agentic AI autonomy expand based on earned trust?. Read that way, a scope record is a snapshot that should get wider over time as trust builds. OpenAI makes a similar argument for its own systems: monitoring and security should scale with model capability Should security controls scale with model capability?. Both treat scope as something that changes over time, not a setting you fix once.

One caveat: the corpus has no studies of how these two layers behave together in shipped products. The pattern library itself admits there is no evidence its controls prevent people from rubber-stamping approvals or other oversight failures. So the mapping is clear on paper, but no one in this collection has tested it.