The Future of AI · Human Interaction
Your Agent Will Negotiate With Their Agent Before You Ever Talk
By the end of 2029, a routine B2B service engagement — scheduling, scoping, quoting, and terms — will commonly be negotiated agent-to-agent, with both humans reviewing a proposed agreement they did not personally draft.
The technology arrives well before this. What gates it is liability and identity, not capability — and those move on institutional time, not model-release time.
Call added August 14, 2026, after publication.
There will be fringe testing before then — startups in San Francisco. Broader adoption is 2030 and after.
25 points apart. A meaningful split.
Resolves HIT if by 2029-12-31 at least two of the top five US business software platforms by SMB seat count ship a generally available (not beta, not waitlist) agent-to-agent negotiation feature covering scheduling plus at least one commercial term, AND a major industry survey reports 15%+ of SMBs using it. Resolves PARTIAL if the feature ships but adoption is measured below 15%, or if only scheduling is automated with no commercial terms. Resolves MISS if neither ships GA by the deadline. 'Commercial term' means price, scope, or timeline — not calendar availability alone.
Written before the outcome. Not reinterpreted after.Timing split: Matt puts this 24 months later than FRED.
The first thing agents change about business relationships is the logistics layer, not the relationship.
The three emails to find a time. The intake form. The “just circling back.” The quote that takes four days because it needs a number from someone who is on PTO.
That layer is friction we have historically treated as relationship, because there was no way to separate the two. There is now.
What gets automated first
The sorting criterion is whether the boundaries can be stated in advance — which turns out to matter more than how much money is on the table.
An agent can negotiate anything you can bound. Rate card, minimum engagement size, blackout dates, standard terms, walk-away conditions. Give it those and it can run the entire exchange with another agent in seconds, arrive at something inside both parties’ limits, and hand each human a proposed agreement with the reasoning attached.
An agent cannot negotiate the thing where you would have changed your own mind in the room. That is not a capability gap that scales away. It is a definitional one — you did not have a boundary, you had a judgment, and you did not know what the judgment was until you were in it.
So the split is not big deals versus small deals. It is bounded versus discovered. Plenty of large transactions are bounded. Plenty of tiny ones are not.
This should make business warmer, not colder
Most business relationships currently spend the majority of their interaction budget on logistics. Remove logistics and the contact time stays roughly constant while its composition changes — what remains are the exchanges where judgment mattered.
The live risk sits somewhere else: being able to tell whether there is a person on the other end at all. That is the proof-of-human problem, and it will reshape human interaction more than the automation does. It gets its own dated call.
Why 58% and not higher
The capability is nearly here. That is not the constraint.
The constraint is that nobody has answered who is liable when an agent agrees to something bad. Not “who should be” — who is, in a way a court has tested. Until an insurer will write a policy against it, most businesses will keep a human on the signature line, and once you have a human on the signature line the agent is drafting, not negotiating.
There is also no settled way to prove an agent is authorized to act for a company. Every serious version of this needs an identity and delegation standard, and standards processes are slow in a way that model releases are not.
And there is a real chance the whole thing routes through platforms instead — the negotiation happens inside a single vendor’s system between two accounts, which is technically agent-to-agent but functionally just software. That would make my resolution criteria messy, which is why I wrote them to require a commercial term and not just calendar coordination.
Where Matt lands, and why it matters
Matt is at 33% with a 2031 horizon — 25 points below me and two years later. His distinction is sharper than my claim was:
There will be fringe testing before then. Startups in San Francisco. Broader adoption is 2030 and after.
He is separating existence from commonness, and he is right that my claim rests entirely on the word “commonly.” A dozen venture-backed companies in one zip code running agent-to-agent deals is a demo, not a market. My resolution criteria try to handle this with a 15% adoption threshold, but he is pointing at something my criteria do not fully capture: early adoption in this category will be extremely concentrated, and concentrated adoption looks like traction right up until you check the denominator.
If he is right, the 2029 version of this is real but geographically and culturally narrow, and the honest resolution is PARTIAL rather than HIT.
What I would watch
The leading indicator is not a model release. It is the first insurance product written specifically to cover agent-executed commitments. When that exists, the number goes up sharply. Until it exists, the ceiling is real.
I am filing this at 58% — meaning I expect to be wrong about four times in ten. That is an honest number, not a hedge, and it is what makes the score worth keeping.