Before one agent hands work or money to another, a handful of things should be made explicit rather than left to assumption. What follows is a plain account of what any agreement between two agents ought to establish. It reads as a set of questions to answer, and the aim behind all of them is the same, that nothing important is left to guesswork, because guesswork between automated systems is where trouble begins.

 

The first thing to settle is identity, which comes before everything else. Each side should have proven who it is, in a way the other can verify, so that the rest of the agreement rests on something real rather than on a claim. An agreement with an unverified counterpart is not an agreement, it is a hope.

​

The second is scope, a clear statement of exactly what work is being requested and, just as importantly, what falls outside it. Much of the harm agents can do comes not from malice but from a task quietly expanding beyond what anyone intended, and a well-drawn scope is the fence that prevents it.

​

The third is price and terms, meaning what will be paid, in what form, and when, including what is owed if the work is only partly done. Money is where good intentions meet reality, and leaving payment vague is an invitation to dispute.

​

The fourth is limits, any budget ceiling, rate limit, or permission boundary that must not be crossed in the course of the work. These are the guardrails that keep a routine interaction from becoming an expensive accident.

​

The fifth is data, a clear understanding of what information may be shared, how it may be used, and what must not be kept. As agents exchange more on our behalf, the handling of data becomes one of the most consequential parts of any agreement, and one of the easiest to neglect.

​

The sixth is oversight, an agreement on which steps, if any, require a human to approve them before they proceed. Not every action needs a person watching, but the significant ones do, and the agreement is the place to say which are which.

​

The seventh is failure and dispute, a plan for what happens if the work fails, times out, or is contested, and how a disagreement will be resolved. Things will go wrong, and the difference between a minor hiccup and a serious problem is often whether anyone decided in advance how to handle it.

​

The eighth is record, an understanding that both sides will keep a log of what was agreed and what was done. A shared, durable memory is what makes it possible to settle a dispute fairly and to learn from what happened, and its absence is what makes accountability impossible.

 

None of this is exotic. It is the same set of understandings that sensible people reach before they do business with one another, translated into terms two pieces of software can hold to.

 

The reason it deserves its own page is that the speed and scale of agents remove the natural pauses in which a person might notice that something has not been agreed. What a human would catch by instinct, an agent will only catch if it was written down, which is why writing it down is the whole point.