One-Way Doors
Agents made every decision fast. That was only ever safe for most of them.
In his 2015 letter to Amazon shareholders, Jeff Bezos split decisions into two kinds. Some are "consequential and irreversible or nearly irreversible," and those "must be made methodically, carefully, slowly, with great deliberation and consultation. If you walk through and don't like what you see on the other side, you can't get back to where you were before." He called these Type 1 decisions, or one-way doors.
Most decisions aren't like that. "If you've made a suboptimal Type 2 decision, you don't have to live with the consequences for that long. You can reopen the door and go back through. Type 2 decisions can and should be made quickly by high judgment individuals or small groups."
The letter's warning was about large companies getting this wrong in one direction: "As organizations get larger, there seems to be a tendency to use the heavy-weight Type 1 decision-making process on most decisions, including many Type 2 decisions. The end result of this is slowness, unthoughtful risk aversion, failure to experiment sufficiently, and consequently diminished invention."
That is the half of the letter everyone quotes. The other half is in a footnote.
The footnote is the part that matters now
Footnote 1 reads: "The opposite situation is less interesting and there is undoubtedly some survivorship bias. Any companies that habitually use the light-weight Type 2 decision-making process to make Type 1 decisions go extinct before they get large."
In 2015 that was the less interesting case, because you rarely saw it. The companies that did it didn't last long enough to be studied, which is the survivorship bias the footnote names.
Autonomous agents make it the interesting case again. An agent running on a schedule has exactly one decision process, and it's the lightweight one. It reads, it decides, it acts, and nobody looks at that particular call. For a two-way door, that's the right process. Bezos would call anything slower a failure to experiment. For a one-way door, it's the process his footnote says companies don't survive.
So the useful question for anyone deploying agents is the one the letter actually asks: which of these calls are one-way doors, and are those running through a different process from the rest?
Reversibility isn't a property of the decision alone
Bezos describes the door as a feature of the decision. With agents, the same decision can be reversible in one place and permanent in another. Three things close a door that would otherwise swing both ways.
It leaves the system. A draft summary is a two-way door. The same summary posted to a shared channel is not. You can delete the message, but you can't make the people who read it unread it, and you can't recall the decision someone made in the ten minutes it was up.
Something builds on it before anyone looks. A product decision written down is easy to change. The same decision after a coding agent has built a feature against it is a merged pull request, a migration, and behavior a customer has already seen. The decision didn't get less reversible. Everything downstream of it did.
Reversing it leaves no record. Changing your mind is only a two-way door if the system knows you went back through. If the old decision stays in place looking exactly as authoritative as the new one, you didn't reopen the door. You added a second one, and the next agent picks whichever it finds first.
That gives a short audit you can run on any agent's actions:
| Action | Door | Why | What it needs |
|---|---|---|---|
| Rewrite an internal record from a source that still holds the truth | Two-way | A wrong copy gets re-read and corrected | Speed |
| Draft output nobody has seen yet | Two-way | Nothing has left the system | Speed |
| Deliver output to people | One-way | Once it's read, it's read | A look before the first delivery |
| Accept a judgment call others will build on | One-way once built on | Work stacks on top of it | A human call before it propagates |
| Replace an earlier decision | Two-way only if marked | An unmarked old decision keeps governing | An explicit record of what replaced what |
Most of what an agent does lives in the first two rows. That's Bezos's main point, and it holds: treat those as one-way doors and you get slow for nothing. Put the caution at exactly the doors, and let everything else run fast.
Where this fits at Brief
Brief's agent design maps onto that table: it finds the one-way doors and puts the checks there, and it records reversals so that changing your mind stays a two-way door.
Delivery is treated as a one-way door. Every new agent in Brief is created paused. When you arm it, Brief runs a preview against real data using the same production executor the scheduled runner uses, and the preview never delivers. You see what the agent would send before it can send anything. Arming a new agent also checks that it can actually reach where it's configured to deliver. A private Slack channel the bot isn't in blocks arming. That's Type 1 treatment, applied to the one step that can't be taken back, before the agent's first walk through it.
Facts move fast, and judgment waits. Brief's agents split along the same line. The ones that deal in observed fact, like Revenue, Pipeline, and Velocity, write directly. A subscription exists or it doesn't, and the agent's job is simply to keep the record true. That's a two-way door, and making someone approve it would be the slowness the letter warns about. The ones that deal in judgment, like Strategy, Positioning, Persona, and Feature, only propose. Each suggestion waits in a review queue until you accept it, edit it, or throw it out. The reason is the table's judgment-call row. What's accepted becomes what a coding agent reads from when it asks Brief what the team has decided, and a judgment call's door closes the moment something gets built on it.
Changing your mind stays a two-way door. The supersession post walks through a team that decides in March that enterprise exports are CSV-only, then decides in June to allow scoped PDF export for admins. In August an agent finds the March decision and builds CSV-only export, perfectly and wrongly. The team did reverse the decision, and even wrote the new one down. Nothing recorded that it replaced the old one. Brief tracks when one decision replaces another, so the March call stays as history and stops governing the next build. That's what makes Bezos's "reopen the door and go back through" true for a team that works through agents. Without it, every reversal leaves the old door open behind you.
Put together, the design runs lightweight where decisions are reversible and heavyweight where they aren't, which is exactly the split the letter asks for. Two of those places get a check before the one-way step, and the third keeps reversals two-way by recording what replaced what.
The line isn't perfect, and it's worth saying where. A preview is a point-in-time check. It shows what the agent produces against today's data, not next quarter's. An empty or failed preview can be acknowledged and the agent armed anyway. And once an agent is live, running it on demand delivers its output. As the dry-run post puts it, there's no way to see what it would say without saying it. The one-way door gets a look before the first walk through, not a lock on every walk after.
Which of your agents' actions are one-way doors, and did anyone look at them before the first time an agent walked through?
Frequently asked questions
What is a one-way door decision? It's Jeff Bezos's term, from his 2015 letter to Amazon shareholders, for a decision that is consequential and irreversible or nearly irreversible. If you walk through and don't like what's on the other side, you can't get back. He called these Type 1 decisions and said they should be made slowly and with deliberation. Two-way doors, or Type 2 decisions, can be reversed, so they should be made quickly.
Why does the one-way door idea matter more for AI agents? An agent running on a schedule makes every decision through the same fast, unreviewed process. That's right for two-way doors. For one-way doors, it's the situation the letter's footnote describes: using the lightweight Type 2 process for Type 1 decisions, which it says companies don't survive long enough to get large.
What makes an agent's action a one-way door? Usually one of three things: the output leaves the system and gets read, something gets built on it before anyone checks, or reversing it leaves no record, so the old version keeps governing. The same decision can be a two-way door in a draft and a one-way door once it's delivered or built on.
Doesn't reviewing one-way doors slow agents down? Only if you review everything. Bezos's main warning runs the other way: the slowness comes from putting two-way doors through the heavy process. Most of what an agent does is reversible and should run fast. The review belongs at the few steps that aren't, and ideally before the first time, not after every run.
How do you make one-way doors safer for agents? Put the check before the point of no return, because afterward is too late. Preview output before it's ever delivered, and keep judgment calls out of what others build on until a person accepts them. Neither of those makes the door two-way: a delivered message has still been read, and work built on a decision is still built. The one step that does keep a door two-way is recording what replaced what when a decision changes, instead of deleting or overwriting the old one, so going back through the door really takes you back.
← Back to Blog