Spend7

What a payment risk score does when an agent has no history yet

AcademySpend7 Content Team9 min read
Ledger rows standing in for the agent payment history that a risk score is built from
Ledger rows standing in for the agent payment history that a risk score is built from Photo via Unsplash.

TL;DR: key takeaways

  • A risk score built on history cannot say much about an agent's first payments. The engine says so, rather than guessing.
  • Below eight settled payments, the relative amount signal is silent. Absolute size still counts from the first payment.
  • In our own run, a $400 payment after three $40 payments scored 11 and was allowed. Nothing about its size was scored at all.
  • For a new agent, the spend cap is the control that does the work. The score is the second opinion.

An agent's first payment is the one a risk score knows least about. There is no record of what it usually spends, which merchants it normally uses or how quickly it normally pays. So the question a developer really asks is simple: what does the score do in that gap, and is an allow decision on day one worth anything?

This post answers it with numbers from the engine itself. We ran the scoring code from our own repository, with a fixed clock and a fixed history, and recorded what each intent scored and why. Where the answer is "not much", we say so, and we show the control that covers the gap.

What the engine can see before an agent has history

Spend7 scores each payment intent as a sum of named signals, clamped to 0 to 100. Three thresholds sit over that score: 50 is a flag, 80 is a deny. Each signal has a ceiling, so no single signal can decide the outcome by itself unless it is a hard cap breach.

For an agent with no history, most behavioural signals switch off. The ones that stay on are:

SignalMaximum pointsWhat it reads with no history
policy_cap_breach100A configured cap the payment would breach. Deterministic, regardless of history.
amount_magnitude22Absolute size, log-scaled from $500 to $50,000. Scored at full weight when there is no baseline.
no_history12Reduces as the agent builds toward eight settled payments.
first_seen_merchant12The merchant has never been paid by this account.
category_prior25A prior from the merchant category, independent of the agent.

Two things are missing from that list. The amount ratio signal, amount_anomaly, needs eight prior payments before it can compare anything. And velocity_burst needs twelve intents in five minutes, which a new agent rarely reaches on its first day.

That's intentional. A median of two or three payments is not a baseline. Scoring a $400 payment as "four times normal" on that basis would be invented precision.

The worked example: one agent, six intents

The agent is research-agent-02. It buys data from api.example.com over x402, with category data_api. We varied its history and the size of the next payment. The clock is fixed at 10 November 2026, 12:00 UTC, the policy lists are empty and there is no graph context, which is the stateless setup a paid-plan account would not hit.

Prior historyIntentDecisionScoreSignals that fired
None$12,000allow42magnitude 15, no history 12, first-seen merchant 12, category 3
None$1,000allow30no history 12, first-seen merchant 12, category 3, magnitude 3
3 × $40$400allow11no history 8, category 3
3 × $40$40allow11no history 8, category 3
8 × $40$400allow20amount anomaly 17, category 3
8 × $40$4,000allow43amount anomaly 30, magnitude 10, category 3

Read the table slowly, because it contains three different failure modes.

Row one. A $12,000 payment from a brand-new agent to a merchant it has never paid is allowed. It scores 42, which is below the flag threshold of 50. The engine has seen a large amount, a new merchant and no history, and it has weighed them up to 42. That is honest, but it is not a block.

Row three. Three prior $40 payments make the agent's typical spend roughly $40. A $400 payment is ten times that, and the engine cannot say so, because it needs eight payments before the ratio means anything. It scores 11, all of it from "we do not know this agent yet". The $400 is also below the $500 magnitude floor, so size contributes nothing. Nobody would flag it.

Row five. Once the agent has eight payments, the ratio signal arrives and the same kind of mistake scores properly: 8 × $40 followed by $400 gives an amount anomaly of 17. But at $4,000, a hundred times the median, the score is 43 and the decision is still allow. The ratio signal saturates at 30 points, and the absolute size adds only 10 more. A hundredfold jump in an agent's spend is a clear signal, and this configuration still lets it through.

None of these results indicate a problem. They are what the scoring says when the evidence is thin, and it is the reason the score cannot be the only control.

Why eight payments

The eight is minHistoryForBaseline in the ruleset: the number of settled payments in the last 30 days before an agent's amount baseline is trusted. The number is a judgement, not a statistical result. Eight is enough to have a median that one unusual payment cannot move much, and few enough that an agent reaches it within a working week if it is busy.

Once that baseline exists the engine does two things. It fires the ratio signal when a payment is at least four times the agent's median, saturating at 25 times. And it damps the absolute magnitude signal to 30 per cent of its weight when the payment sits inside the agent's normal range. The second behaviour matters to a business that routinely moves six figures: it should not be told every payment is large.

The engine also records the shortfall in plain words. A no_history signal reads "N settled payment(s) ... against the 8 needed for an amount baseline. Until there is one, an unusual amount cannot be distinguished from a normal one, so this scores as mild uncertainty rather than as safe." That sentence is the design. Missing history is not treated as a clean record.

Caps are what catch a new agent

For a new agent, the control that does the work is a spend cap, not the score. Here is the same $12,000 payment with a per-transaction cap of $1,000 on research-agent-02:

  • Decision: deny, score 100.
  • Signals: policy_cap_breach 100, with the magnitude, no-history, first-seen and category signals still listed underneath.

A breached cap adds 100 points, which clears the deny threshold on its own. The cap does not depend on how much the engine knows, which is exactly the property a new agent needs. You do not have to wait for eight payments to say that $1,000 is the most the agent may spend in one go.

We recommend setting two caps before an agent's first live payment:

  1. A per-transaction cap at the largest single payment you would accept from the agent without a human. This is deterministic.
  2. A rolling daily cap at roughly twice what you expect the agent to spend. The point is to catch runaways. Rolling means the last 24 hours, not since midnight, so a cap cannot be doubled across a date boundary. The x402 spend limits guide walks through the setup step by step.

Once the agent has a history, you can tighten these caps to what it actually does. Until then, they are the only controls that do not depend on what the engine has learned.

What the research says about thin history

The general problem is old. Fraud teams call it the cold start: a new account or a new device has no behaviour to compare, and most models are weakest exactly where the first decision is made. The common responses in the literature and in vendor material are the same three, and they transfer to agents with little change.

First, use signals that do not need history: identity and device checks at onboarding, and third-party data where it exists. Microblink's guidance on new account fraud argues that risk teams cannot rely on account-level anomaly detection at onboarding and should lean on front-end verification. For an agent, the equivalent is verifying the agent's identity and its mandate before it spends, rather than after.

Second, train on small labelled sets where they exist. Amazon Fraud Detector's Cold Start feature describes training from a minimum of 100 events, at least 50 of them classified as fraud. Few agent deployments can meet that bar, which is why an explicit ruleset with named weights is easier to defend than a model trained on thin data.

Third, borrow signal from connected entities. The ColdGuess paper on e-commerce listings passes information from existing sellers to new ones, and reports large gains in the extreme cold-start case. Spend7 does this on paid plans through the cross-agent and cross-merchant graph, which is why the graph is reported as unavailable on free accounts rather than silently absent.

None of these sources measures agent payments directly. Treat them as the shape of the problem, not as benchmarks for our numbers.

Common mistakes with a new agent

Treating allow as clearance. An allow on thin history means the engine had little to compare, not that the payment matched a known pattern. Log the no-history signal and review it.

Waiting for the baseline before setting caps. The caps are what protect the first week. The baseline is what tightens them later.

Using one global cap for a fleet. A global cap tells you the fleet overspent. It does not tell you which new agent did it. Give each agent its own cap, with a global one on top, and the tightest binds.

Ignoring the merchant. A new agent paying a merchant it has never used is a signal in its own right, and it is scored that way. Pair it with a merchant allowlist for anything that matters.

Reading the decision without the signals. The headline decision hides how close a payment came. A 42 and a 49 are both allows, and the difference is worth knowing.

Where to go next

If you are setting up a new agent now, start with the caps and the API reference, which includes the context object for scoring a policy against sample history before you enable it. The engine's behaviour on risky payments is covered in transaction monitoring for agent payments, and the wider picture of where checks belong is in agentic payments: what they are and where checks belong.

For what a whole review process looks like once an agent has been live for a while, autonomous agent payment monitoring is the next read. And if you are still choosing a rail, the rails comparison sets out which payment protocols expose the signals discussed here.

Method and limits

The figures in the worked example come from running Spend7's scoring engine at ruleset version 2026.08.2 on the sample intents above, with a fixed evaluation time and no graph context. They are the engine's output for those inputs, not a sample of live traffic, and they will change if the ruleset changes. The version and its hash are returned with every real decision, so a figure quoted from one payment can be checked against the ruleset that produced it.

Nothing here is a guarantee of fraud prevention. The engine produces a pre-payment decision and a set of reasons. Moving the money, and the decision to proceed, remain with the system that calls it.

Frequently asked questions

What does a payment risk score do for an agent with no history?
It still scores the payment, but only on what it can see without a baseline: the absolute size of the amount, whether the merchant is new to the agent, the category prior, and the missing history itself. Each of these is shown as a named signal with its points. It does not assume the payment is normal just because nothing is known against it.
How many payments does an agent need before its amount is judged against its own history?
In the current ruleset, eight settled payments in the last 30 days. Below that the relative amount signal stays silent and the engine reports the shortfall as a no-history signal, weighted by how far the agent is from the eight it needs.
Why is a $12,000 payment from a brand-new agent allowed?
Because on the scoring alone it totals 42, under the flag threshold of 50. The score is not a verdict on the payment. It means the behavioural evidence was thin, so a spend cap is what should stop it. A per-transaction cap of $1,000 denies the same payment outright, because a breached cap adds 100 points.
Should I treat an allow decision for a new agent as safe?
No. An allow on thin history means the engine had little to compare against, not that the payment matched a known pattern. Set a per-transaction cap and a rolling cap for every new agent, and read the no-history signal in the log rather than the headline decision.

Score a payment before it settles

Spend7 returns allow, flag or deny in one call, with the signals that produced it. The free tier covers a single agent, its spend caps and its full decision log.

Keep reading