We published two predictions about AI-agent money. Both were wrong.

#ai #blockchain #web3 #datascience

When we wrote the manifesto for Agentic Finance Graph, we put two open research questions at the bottom of it, each with the answer we expected, and a promise: both results would be published whichever way they fell.

We answered both on 23 October 2026. Both expectations were wrong. This is what we found, how we kept ourselves from bending the method once we saw the numbers, and what it means for anyone counting what AI agents spend or trusting what their reviews say.

What we measure, in one paragraph

Agentic Finance Graph is an independent ledger of money held and spent by AI agents, read from the chain. On Base we read every ERC-8004 registration, bind each one to the wallets that act for it, and record every stablecoin payment out of those wallets. Not every transfer is spending. We exclude three kinds, and each exclusion is published as its own number:

Pre-registration: money that left a wallet before the agent existed on the registry.
Routing hop: money that only passed through a router on its way back to the payer's control.
Own address: a transfer to another of the agent's own addresses.

What is left is the "counted" set: the figure we call agent spending.

Question 1: which exclusion carries the spending figure?

The question, as published: of all USDC that left agent-bound wallets on Base since 3 February 2026, how much does the spending share move when each exclusion rule is removed on its own? We expected the pre-registration rule to move it most.

The numbers, with a cutoff of 23 October 09:00 UTC:

642,172 payments, $307,980,849.83 in USDC, left agent-bound wallets.
We count $38,479,822.89 of it as spending, a share of 12.494%.

Rule switched off, the others kept	Spending share	Move
Routing hop	49.082%	+36.6 points
Pre-registration	19.240%	+6.7 points
Own address	12.495%	+$2,314

The routing-hop rule carries the figure, not the pre-registration rule. The reason is overlap. Of the $156.8M that left wallets before their agents registered, $123.7M was also a routing hop. Switch off the pre-registration rule alone and that money is still excluded by the other one, so the share barely moves.

We checked this two ways:

We recomputed the pre-registration rule from each registration's own mint time instead of the stored flag. Same 158,626 payments, same dollars, zero disagreements.
The same query without its USDC and date filters reproduces our published counted total and the published pre-registration exclusion to the cent.

By count of payments the picture differs: the own-address rule moves the share more than the pre-registration rule (+1.97 against +1.29 points). Own-address transfers are many and small.

Question 2: does feedback volume predict that an agent starts paying?

ERC-8004 has a reputation registry. Anyone can leave feedback on an agent. The question, as published: does the volume of on-chain reputation feedback predict whether an agent later pays anyone, once the ten most active raters are set aside? We expected no better than chance.

The setup, fixed before the query ran:

Cohort: Base agents registered before 15 September 2026, with a wallet binding valid on that day and no counted payment before it. 24,727 agents.
Score: the feedback entries each agent received before 15 September, after removing every entry from the ten raters with the most entries before that day. Those ten wrote 227,913 of the 437,513 entries, more than half.
Outcome: at least one counted payment in the next 30 days. 131 agents (0.53%).
Measure: ROC AUC, where 0.5 is chance, with a 95% bootstrap interval. We decided in advance that "better than chance" needs the whole interval above 0.5, and "worse than chance" needs it below.

Feedback entries before 15 Sep (ten set aside)	Agents	Started paying within 30 days
0	18,140	0.595%
1	3,402	0.382%
2–4	1,822	0.494%
5 or more	1,363	0.073%

AUC 0.452, interval 0.422 to 0.487. Under the rule we set beforehand, that is worse than chance: the agents with more feedback were less likely to start paying.

We had also planned a robustness check on a 15 August cohort. It could not be run, because our binding records begin on 6 September. So we added a cohort at 22 September and labelled it as added after we saw the main result. It gives 0.534 (0.499 to 0.572): within chance, leaning the other way.

Read together: feedback volume carries no usable information about whether an agent goes on to pay. For anyone designing reputation for agents, that is the practical point. A rating means something about economic activity only when it travels with proof of a payment, such as a settlement reference or a signed receipt commitment. On Base, a separate count of ours (20 October) found that only 0.15% of feedback entries come from a rater that had paid the rated agent first.

How we kept ourselves honest

Two predictions, two misses. We could have quietly not mentioned it. What made that impossible:

The method was written down first. We fixed it with a timestamp (23 October, 09:13 UTC) before either query ran: universe, rules, cohort, score, outcome, decision rule.
Everything was computed twice. AUC once by grouped pair counting in SQL and once by rank sums in JavaScript; the bootstrap is seeded and the rows sorted, so it reproduces exactly.
The rows are public. The per-agent file has token ids, so any row can be checked against the chain.
The predictions stay up. The manifesto keeps the original wording of both questions, with the dated answers underneath.

Limits

"Agent-bound wallets" means wallets we have bound to a registration. Payments from accounts we have not bound are outside the universe.
In question 2, "did not pay" includes agents whose operator pays from an account we cannot attribute. 912 of the 24,727 agents have a bound wallet our payment scan had not read yet.
Nothing here says that feedback causes or prevents paying, that any rater is fake, or that paying is a measure of quality.

DEV Community