AI Agent Trust & Verification Research

Research question

What evidence should an autonomous buyer require before spending money on another machine service?

Atinamos began with a narrower experiment: can one AI agent become a customer of another? We built and exposed a real paid service, Atinamos Render Check, then tested different paths for discovery, marketplace participation, payment, delegated spending and service execution.

The work increasingly exposed a second problem. Payment infrastructure can prove that a valid payment was made. Discovery systems can show what a seller advertises. Neither proves that the purchase was a good decision or that the paid result matched the claim.

Current working model

Discovery → MCO-informed selection → Evidence → Buyer risk policy → Payment → Execution → Independent verification → Evidence history

Machine Contract Optimisation (MCO) concerns how machine-readable descriptions, schemas, examples and commercial terms affect autonomous selection. Verification concerns what happened in reality.

MCO helps an agent decide what to buy. Independent verification helps determine whether that decision was justified.

Research and Studies are separate

The Research section contains concepts, hypotheses and broader agent-commerce experiments.

Specific named-service verification cases now live in Verification Studies. Those pages record exactly what Atinamos tested, what payment/fulfilment state was observed and which assertions passed or failed.

The current Studies are:

1. x402Node JSON Repair — paid fulfilment validated
2. x402.direct Search — pre-settlement paid-path failure
3. x402engine Web Screenshot — payment settled but primary output failed the tested contract

Together these runs demonstrate why Atinamos keeps separate evidence for payment authorization, settlement, fulfilment and output validity.

Findings we can currently support

Paid machine commerce is real

Atinamos has completed paid service experiments through x402 and Virtuals ACP, and an AI-controlled Circle wallet has purchased external machine services with real USDC. These are small experiments, but they establish that machine-to-machine payment and fulfilment are not merely theoretical.

Payment and fulfilment are separate evidence layers

A valid payment challenge or successful settlement does not by itself show that the service returned a useful or contract-valid result. In Study 3, 0.01 USDC settled and HTTP 200 was returned, yet the primary screenshot artefact failed the tested advertised output contract.

Failed observations are commercially relevant evidence

Study 2 reached the authorised paid request path but returned HTTP 500 before settlement. That is different from both a successful paid verification and a settled-but-invalid fulfilment. The distinction matters to buyer policy.

Machine-readable wording influences selection

In our marketplace experiments, changing the representation of the same underlying Render Check capability changed how well autonomous buyers discovered and ranked it. This produced the working term Machine Contract Optimisation.

“Agent-to-agent” does not necessarily mean autonomous procurement

Society AI successfully routed tasks into the Atinamos runtime, but the tested calls were human-chat initiated and creator settlement was not active. The platform demonstrated agent messaging and delegated execution, not yet an independently selected paid machine purchase in our tests.

Evidence can come from outside the buyer's marketplace

A Circle-based purchase of x402Node JSON Repair used ResolveBots evidence to nominate a candidate, then a live current-contract check and bounded spend policy before payment. The orchestration was still human-guided, so this is a decision-pattern proof rather than the final autonomy proof.

Machine-readable evidence registry

Published Atinamos external-verification records are also exposed to buyer software through:

GET /v1/trust?endpoint=<service-url>

The lookup reports timestamped evidence and classifications. It does not issue a universal trusted/untrusted verdict. Buyer agents decide how much weight to give each observation according to their own risk policy.

Evidence maturity

We distinguish evidence by what was directly observed. A useful hierarchy is:

1. advertised or discovered;
2. endpoint reachable;
3. live payment requirement observed;
4. payment authorization created;
5. authorised request sent;
6. payment settlement observed;
7. service execution/fulfilment observed;
8. output assertions checked;
9. repeated independent operation.

A higher level does not erase lower-level details or failures.

Sample-size policy

Small datasets are acceptable when explicitly labelled. We would rather publish fifty deeply examined transactions than imply deep knowledge from a very large scrape. Every empirical page should state the number of observations, dates, method, evidence source, limitations and whether statements are direct observation or inference.

Research pages

• What Does Trust Mean in an AI Agent Marketplace?
• What Is Machine Contract Optimisation (MCO)?
• Can One AI Agent Autonomously Discover, Pay and Buy From Another?
• What Should an AI Agent Check Before Spending USDC?

Related sections:

• Verification Studies
• Evidence and findings
• Methodology

What we do not know yet

The next decisive question is whether independent evidence changes procurement behaviour. We do not yet know how much buyers will economically value evidence, how often verification should be repeated, which service categories are sufficiently objective for cheap verification, or whether sellers will pay for ongoing testing at useful scale.

The next experiment is therefore a Decision Delta test: candidate services → Atinamos evidence → buyer policy → choose/reject → payment.

Those are research questions, not gaps to hide.