As we think about handing more work off to AI agents, we need a framework for understanding both the type of work we're delegating and the consequences when an agent gets something wrong.
At Mostly Serious, we've long used Amazon and Jeff Bezos's framework of one-way and two-way door decisions to understand how much time and scrutiny a decision deserves.
A one-way door decision is difficult, costly or impossible to reverse once it's made. These decisions deserve more consideration, debate and often executive-level buy-in. A two-way door decision is easy to reverse. You can move quickly, experiment and make a different decision if the first one doesn't work.
The original framework helps people decide how carefully to approach a decision. We can apply the same framework before delegating an action to an agent. The question stays roughly the same, but we apply it to the action being taken.
If the agent gets this action wrong, can we undo the consequences?
"Action" is important, because within the same workflow, the agent may be working with both kinds of doors.
For example, having an agent research a prospective lead and prepare an internal brief for the salesperson is a two-way door with relatively low consequences. If the research is bad, a person can catch it during their review and disregard it or do their own research.
Another example is an agent drafting a proposal. The draft can be reviewed, revised, and updated by a human before it's ever sent. However, sending that same proposal is a different type of action and gets much closer to a one-way door. Proposals typically include pricing, define the team's approach to the work, and make commitments on the company's behalf. Once it goes out the door, those details may be difficult or expensive to undo.
Speed and scale are also considerations when working with agents. A single update to the CRM is pretty easy to reverse, but thousands of incorrect updates made without oversight, meaning they may be missed for quite some time, can cause significant problems. So we have to think about the scale of the decision made by agents, not just whether it can be reversed on the basis of an individual action.
The one-way versus two-way door distinction gives us a useful mental model for what kind of work agents can do. It helps us determine the level of autonomy they should have. Two-way door actions can often be delegated with more autonomy and less oversight. As an action becomes harder to reverse or greater in scale, the system needs stronger guardrails and explicit human authority before the agent can act.
Interactive framework
Walk through the Agent Door Test
Start with one exact action. Then test its consequences, scale, and authority before deciding who acts.
- Action
- Reversibility
- Speed + scale
- Authority
- Decision
Importantly, giving agents authority doesn't mean that a human has to be in the loop every single time. It means that humans lead by defining the boundaries before the agent acts. For example, a company might authorize an agent to issue refunds within a threshold that varies by product type, with daily, weekly, and monthly limits to control total spending. Anything beyond an applicable limit requires a human decision before the refund is issued.
That kind of bounded authority lets the agent move quickly without removing human control. But it doesn't change the nature of the door. Testing and monitoring can give us confidence that the agent will behave reliably, but they still don't make irreversible consequences reversible.
Which brings us back to the question we should ask before an agent acts:
- What specific action is the agent taking?
- How easy is it to undo the consequences?
- What happens if the action is repeated at scale?
- Is the action within the authority a person has explicitly defined?
That is what we're calling the Agent Door Test.