What are one-way and two-way doors for AI agents?
- One-way and two-way doors for AI agents
- One-way and two-way doors sort an AI agent's actions by how easy a mistake is to undo. Two-way doors, like moving a meeting, can be handed to the agent quickly. One-way doors, like buying a ticket you can't refund, need much more evidence, and some should always stop for a person to confirm.
How does the door change what an agent can do?
How easy a mistake is to undo matters more than how big the action looks. Moving a meeting is a two-way door. If the agent gets it wrong you move it back. Buying a ticket feels like the same size of job, but getting out of it means reselling, and that costs fees, time and often money.
Most actions fall into one of three kinds, and the kind sets four rules for that action's ladder in earned autonomy.
| Two-way door | Costly to undo | One-way door | |
|---|---|---|---|
| Highest rung | 5 | 3 or 4 | 2, or 3 if the user asks for it |
| Evidence to climb a rung | A few clean yeses | Many yeses with few edits | Lots, and the user asking for it |
| What one mistake costs | Down one rung | Down two | Back to suggest |
| What counts as unusual | Very different from normal | Somewhat different | Anything new |
The ladder has five rungs, from suggest to act and escalate.
What does it look like in practice?
| Action | Door | How far it can climb |
|---|---|---|
| Move or delete a calendar event | Two-way | All the way, quickly |
| Queue songs, build a private playlist | Two-way | The top, almost straight away |
| Buy a ticket you can return | Costly to undo | Act and tell, under your usual price |
| Buy a ticket you can't refund | One-way | Prepare, or act and tell for nights you always go to |
| Send a gift | One-way | Prepare, for a long time |
| Send an email to your boss | One-way | Prepare, unless you've asked it to |
| Buy tickets for friends | One-way, other people's money | Prepare |
| Share a track or post as you | One-way, your reputation | Suggest |
So the same agent can be very autonomous about listening and careful about buying, for the same person, at the same time. Nobody has to pick one autonomy setting for the whole product. The music agent post works this through in full.
Can you turn a one-way door into a two-way one?
Often, and it's usually the best design move. Pencil in the date and say how long until it sells out, instead of buying. Wait a minute before sending the email, with a cancel button. Pick the refundable option when it costs a little more. Each of those lets the agent do more on its own, because a mistake is cheap again.
The platform can change the door too. Where tickets can be returned for a full refund, buying stops being a one-way door, and the agent can safely do more.
Where does the idea come from?
Jeff Bezos split decisions into two kinds in his 2015 letter to Amazon shareholders. One-way doors are decisions that are "consequential and irreversible or nearly irreversible". Two-way doors are ones you can walk back through. I applied it to AI agents in Earned autonomy, where the door decides how high each action's ladder goes, how much evidence it takes to climb and how far one mistake drops it.
Related ideas
- Earned autonomy. How an agent earns the right to act on its own, one person and one action at a time. In earned autonomy.
- The agent scope doc. One page before the PRD on what an agent can see, do and must never do. In the agent scope doc.
- Gig Agent. A working prototype where each action climbs its own ladder. At labs.mattalcock.com/gig-agent.
FAQ
What are one-way and two-way doors for AI agents? One-way and two-way doors sort an AI agent's actions by how easy a mistake is to undo. Two-way doors, like moving a meeting, can be handed to the agent quickly. One-way doors, like buying a ticket you can't refund, need much more evidence, and some should always stop for a person to confirm.
Who came up with one-way and two-way doors? Jeff Bezos, in his 2015 letter to Amazon shareholders, about company decisions. Applying it per action to AI agents, as part of earned autonomy, is from Earned autonomy.
What should an AI agent never do without asking? Anything that's a one-way door it hasn't earned. Spending money you can't get back, sending messages as you, posting in public and spending other people's money should need your confirmation until the agent has a long track record, and some should always ask.
Why does undo matter more than how big an action is? Because a mistake on a two-way door costs a minute to fix, and a mistake on a one-way door can't be fixed at all. Moving a meeting and buying a ticket look like the same size of job, but only one of them can be walked back.
Should an action that involves other people be treated as a one-way door? Yes. Buying tickets for friends spends their money, and sharing a track speaks for you in public. Those stay low on the ladder however good the agent's track record is with you, because the person who'd pay for a mistake never said yes to the agent.
To cite this, use “One-way and two-way doors for AI agents”, Matt Alcock, mattalcock.com/ideas/one-way-two-way-doors.