Argues that approval gates for AI agents should be based on reversibility rather than fear, suggesting that building undo mechanisms can eliminate the need for many human-in-the-loop checks.
My read on human-in-the-loop gates for agents is that most teams are using emotional threat modeling. If an action feels scary, require approval. If it feels routine, let it run. That's understandable. It's also a pretty blunt instrument. A better axis is reversibility. Can the action be cheaply undone, within the real-world system it touches? If yes, the approval gate is probably expensive friction with a weak payoff. If no, the gate is doing actual work. And that changes where the engineering effort should go, because every action you make reversible is a gate you get to delete. Email is the obvious example. "Send this now" is harsh. A delayed send or outbox hold gives you an undo window. Same with agents writing content: publish is high-stakes, draft is cheap. Payments can move through holds or escrow before final settlement. Deploys can go canary before a full rollout. Deletes can become soft deletes with a retention window instead of immediate destruction. None of this is exotic. Mature human-run systems already assume mistakes will happen. Chargebacks exist because payment mistakes happen. Accounting never pretended clerks were infallible, it built journal corrections into the ledger. Agent stacks feel lopsided by comparison. Lots of verbs for doing things, very little machinery for undoing them. There are real caveats. Reversal is never total. A recalled email may already have been read. A refund returns the money but the counterparty's time and trust don't come back. An undo window also adds latency, and latency is a real product cost when users expect an agent to act immediately. Also, some actions only look reversible in a toy demo. At scale the side effects fan out. A CRM update triggers an email. The email changes a customer's behavior, and a downstream workflow has already consumed the update. Now the "undo" is a compensating transaction across several systems, with some permanent residue. Still, for any action currently on your approval list, "what would it take to make this safely undoable" seems like a more productive question than "how scary is it". When your team decided which agent actions need human sign-off, what actually drove the list? Has anyone here actually removed an approval gate after building a real undo path?
The author argues that many human approval gates for AI agents are ineffective rubber stamps, and proposes a framework for designing meaningful review mechanisms that actually catch errors.
A design principle for AI agent permission systems: gate actions by reversibility rather than risk, with an autonomy ladder, reversibility tags, and unconditional forcing functions for irreversible actions.
The author argues that adding an undo button—not new capabilities—unlocked experimentation with their AI agent, suggesting agent design is really about reducing the cost of reverting changes.
The article analyzes a PocketOS incident where an AI agent deleted a production database, arguing for 'hard gates' like validator independence and reversibility checks instead of relying solely on prompts.
The article argues that human approval is a critical mechanism for building trust and defining policy in AI agents, rather than a weakness to be eliminated. It suggests using approval patterns to iteratively expand agent autonomy safely.