← Glossary

The commit problem

The commit problem is deciding when an AI agent may take an irreversible or intent-disclosing action before it is fully certain.

Definition

To talk at human speed, an agent has to start working before it knows exactly what’s needed. That’s fine when the work is a database read it can throw away. It becomes a trap the moment the work is placing a call, sending a text, or canceling a booking — actions it can’t take back. The commit problem is the policy for which actions may be taken, when, and which must never be taken speculatively.

It runs along two axes, not one. State-reversibility asks whether the effect can be undone. Intent-disclosure asks whether merely issuing the action leaks what the agent is about to do to a third party. A “read-only” availability check can be perfectly reversible in state yet impossible to un-disclose.

Why it matters

For an agent that acts over the phone, the commit problem is what makes it trustworthy. The set of actions it can’t take back is short and should be fixed and declared outside the model, because a model’s in-the-moment judgment about whether an action is safe to commit is exactly the judgment you can’t fully trust.

Getting it right is what separates an agent that sounds capable from one you’d let act in your name.

The two axes of an action

State-reversibility
  • Can the effect be undone?
  • A database write can often be rolled back
  • A placed call, sent text, or left voicemail cannot
  • Compensation (“call back and apologize”) is not rollback
Intent-disclosure
  • Does issuing it leak what you’re about to do?
  • A vendor’s logs hold the query whether or not you proceed
  • A read can be costless to undo yet impossible to un-disclose
  • Defer intent-leaking lookups until after the okay

Examples

  • Reversible: check availability, draft a message, look up a record
  • Irreversible: place a call, send a text, leave a voicemail, cancel a booking
  • Intent-disclosing: ask a vendor about a slot before the owner has decided

Safety & disclosure

  • Type the tools, don’t ask the model: each action gets a declared effect on both axes.
  • Make the commit point late and revocable — cancelable until dispatch, so “wait, no” is honored to the last instant.
  • Confirm by narrating the action (“calling them now to move it to 7”), with barge-in as the abort.

FAQ

Why can’t the model just decide what’s safe to commit?

Research shows models that correctly refuse a harmful request in text will still attempt the forbidden tool call. So the set of irreversible actions has to be small, fixed, and external to the model rather than left to in-the-moment judgment.

Isn’t a read-only action always safe to run early?

No. A read can be irreversible in the privacy domain even when it’s reversible in the state domain — issuing it can disclose the owner’s intent to a third party before the owner has committed. That’s the intent-disclosure axis.

See it in the product: how Call My Agent works · pricing.

Sources
Last updated June 2026