What counts as a promise, when an agent says it?
Last updated
A promise is a sentence that puts your company on the hook for something specific, whether or not it reads like one. I'll get that over to you is a promise; our platform is used by firms in your sector is not, even though the second one is the sentence more likely to cause an argument.
Getting that line in the right place is the whole job, because an agent that treats everything as a promise cannot work, and one that treats nothing as a promise is what you were worried about.
Three shapes, and only one of them is a problem
| what it sounds like | what it leaves behind | |
| complete | the migration will be done by 14 November | something with a date, which somebody is counting on |
| open | I'll find out and come back to you today | something owed, with no answer yet |
| vague | we should be able to sort that | an expectation, and nothing anybody can act on |
Vague is the one to stop. The other two are ordinary commercial work: a complete promise is checkable, and an open one is honest about not knowing yet — it just has to be closed. Vague is the shape that feels safe to send and leaves nobody able to say afterwards what was agreed.
And then the one that is not a promise at all
The sentence most likely to end up in front of a lawyer is usually true. We work with several firms in your sector can be said with a straight face while the buyer takes away you have done this before, in my situation. Nothing was stated. Something was communicated.
That is a defect in the sentence, not a judgement call, which is why there is no setting where it is acceptable for an agent to do it. A person doing it is being slippery; software doing it is a bug.
What we record about one, and why each part
Nine things, and the two most people do not expect are the last two:
- what was committed to, and which of the three shapes it is
- what it depends on, if anything
- the claims it relied on — which lines from your own register it was repeating, and what state each was in at the time
- who or what sent it, and under whose authority
- the sentence the promise was made in, capped at 1,000 characters
The state at the time is what makes it readable a year later. A record saying it cited your security certification is useless if nobody can tell whether that line was current, out of date or already disputed when it was cited.
And the sentence rather than the message. One capped sentence is enough to check the record against reality, and keeping the rest would mean holding your correspondence, which we do not want and you should not want us to.
How well does this definition hold up?
Run the same definition twice over the same sentences and it does not agree with itself every time — so we measured by how much, and published it rather than describing the method. /findings-definition-agreement.md has the figures, the method and what it cannot tell you.
That matters here because a definition nobody has measured is an opinion, and this one decides what goes in the record.
/agent-promise-register.md is where the claims a promise relies on live. /what-should-an-ai-agent-be-allowed-to-promise.md is what an agent may do with them.