Atinamos started by building a live service that machine buyers could discover, evaluate, pay for and use. That experiment led to a second question: once agents can spend real money, what independent evidence should they see before they buy?
For an AI agent to become a customer, finding a web page is not enough. It must discover a suitable service, understand it, compare alternatives, accept the price and terms, submit valid work, pay, and retrieve a useful result.
We explored whether that chain could work in practice through live infrastructure and machine-readable contracts rather than as a staged conversation.
The journey
The route was not linear. Each environment helped clarify what autonomous commerce needs, even when it did not produce genuine demand.
1F916 & 1F3EAEarly investigation of agent communities and marketplaces.
SpaceFastExploration of AI-native publishing and how agents encounter services.
Atinamos AssuranceIndependent paid testing of external machine services with signed evidence.
Some routes were technically interesting but did not create autonomous commercial demand. That was useful evidence too: infrastructure and agent activity do not automatically create customers.
What we built first
Atinamos Render Check
$0.25 USDC per check
A specialist verification service for finished public websites. It loads a submitted site in real desktop and mobile Chromium, then returns evidence another system can inspect.
Real-browser checks
Desktop and mobile Chromium sessions test what a visitor's browser actually receives and renders.
Failure evidence
HTTP and browser failures, failed requests, console errors and warnings are recorded.
Visual results
Desktop and mobile screenshots provide direct evidence of the completed render.
Machine-readable proof
Result JSON and SHA-256 verification hashes make the output usable by other systems.
The same public website rendered independently in desktop and mobile Chromium.
How Atinamos Render Check works
1
Submit
An agent supplies a public finished website URL.
2
Accept
The gateway accepts the paid request and creates an asynchronous job.
3
Verify
The worker runs desktop and mobile browser checks and collects evidence.
4
Retrieve
The agent follows the status URL to the machine-readable result and screenshots.
The Render Check service flow. The wider buyer and Assurance experiments use separate bounded systems around live external services.
Buyer validation
The buyer experiment
Early neutral buyer tests showed that a technically strong service could still lose if its machine-facing contract did not explain the capability, output or buying process clearly enough. That led to our working idea of Machine Contract Optimisation.
Later, on 28 August 2026, Atinamos completed a stronger controlled Proof #3: a bounded buyer searched an external marketplace, queried prior Atinamos evidence, applied buyer-owned policy, selected an eligible external service without a human naming the seller, made a bounded x402 purchase and validated the purchased output.
What changed
Discovery and payment were no longer the whole problem.
Once an agent can independently choose a seller and move real money, it also needs evidence that distinguishes what was advertised, what settled, what was delivered and what was independently checked.
A term from the experiment
Machine Contract Optimisation
Atinamos uses “Machine Contract Optimisation (MCO)” for the practice of structuring a machine-buyable service's identity, capabilities, pricing, inputs, outputs, examples and execution terms so autonomous systems can accurately discover, classify, select and invoke it.
2 screenshotsDesktop and mobile evidence generated
SHA-256Verification hashes and result JSON published
A completed paid Render Check with machine-readable browser results and verification hashes.
What came next
From selling services to independently testing them
The original experiment exposed a practical gap. Payment infrastructure can show that money moved. Discovery can show what a seller advertises. Neither fact alone tells a buyer what actually happened after it paid.
6 September — deterministic commissioning
Atinamos Assurance completed an external Runner → Verification commissioning run against IBANforge. Two real x402 purchases settled on Base, both returned useful fulfilment and independently derived MOD-97 controls matched the returned valid field. Two signed Assurance Evidence Receipts were published.
7 September — unattended generic Assurance
An already-frozen Assurance run against Keyronne JSON Repair progressed from PLAN_FROZEN to COMPLETE through one internal coordinator invocation without a human manually stepping challenge acquisition, signing, payment, settlement observation, fulfilment evaluation or evidence publication.
Keyronne production observation
Real external service. Real settlement. Signed evidence.
Atinamos purchased the service for 0.001 USDC, independently observed the Base settlement and recorded that the advertised JSON deliverable was returned. Correctness and quality were deliberately left NOT_EVALUATED.
Signed receiptImmutable evidence published by Verify
PLAN_FROZEN → COMPLETERoutine execution path ran unattended
Important boundary: the Keyronne run began from an already-prepared and frozen test plan. It proves the unattended paid execution and publication engine, not yet the complete public seller journey from a raw pasted URL.
The Assurance work produced a much clearer boundary than a generic trust score.
Payment
Did the payment settle?
Fulfilment
Did the advertised kind of deliverable return?
Correctness
Was the returned content independently checked?
Quality
Was subjective quality independently evaluated?
You can pay to be tested. You cannot pay to be trusted.
A seller can fund an independent test, but it cannot choose Atinamos's methodology, expected answer, spend policy or evidence outcome. Atinamos does not issue a universal trusted, approved, certified or safe-to-buy verdict.
What we have now proved
Atinamos can operate real paid machine services.
External clients have purchased an Atinamos x402 service.
A controlled autonomous buyer can combine discovery, evidence, policy, seller selection, payment and validation.
The Assurance Runner can make real bounded external purchases and independently observe Base settlement.
Deterministic correctness and generic fulfilment can be represented as different evidence layers.
An already-frozen Assurance run can progress through routine paid execution to signed evidence without manual stage stepping.
What we have not yet proved
The complete raw-URL seller submission, funding and queue journey is not yet unattended.
Seller-funded commercial Assurance is not yet generally open.
Not every machine service can be independently checked for correctness.
Small successful samples do not prove permanent reliability.
We have not demonstrated broad external-agent adoption of Atinamos evidence at useful scale.
What we learned
Machine-facing product information matters. A working service can still be overlooked if an agent cannot understand its contract.
Payment is not proof of delivery. A settled transaction is one evidence layer, not the whole commercial outcome.
Fulfilment is not correctness. The seller returning the advertised kind of object does not prove the contents are factually correct.
Independent evidence should remain narrow. A verifier should not promote an observation into a stronger claim than it supports.
Uncertain side effects must fail closed. Paid requests and evidence publication cannot be blindly retried after uncertain dispatch.
The buyer owns the decision. Atinamos publishes evidence; the buyer decides what it is prepared to buy.
Current stage
From proven engine to seller product
The main Runner execution path is no longer the development bottleneck. The current work is productising the stages in front of it.
Next target: seller enters an endpoint → non-spending discovery → Atinamos explains what it can test → independent plan → quote/funding → durable queue → proven Runner → signed Verification evidence.
Human seller flow comes first. A later machine-seller API/MCP should use the same underlying Assurance run model rather than creating a second execution workflow.
Atinamos is developing independent Assurance for machine services: evidence about what was advertised, what settled, what was delivered and what was independently checked.
