x402-winning Verun charges AI agents 0.01 USDC per trust verdict on Algorand

When an agent wants to move half a million euros

The demo opens on a familiar device: an AI agent named agt_openclaw_01 shopping for €500,000¹ of SpaceX pre-IPO shares. It scans European investment platforms, filters by ‘MiFID II + AI Act audit trail,’ matches tokenforge² with a trust score of 920 out of 1000, and recommends the order to a human, who approves it in one click. Verun³, the project behind that scene, is selling the question at its center: as AI agents begin executing token transfers, minting securities and placing investment orders on their own, there is no widely adopted standard for who an agent is, whether it can be trusted, or whether it should be allowed to act in a regulated market.

Verun's answer, prototyped on Algorand Testnet by BCP Partners GmbH⁴, a Berlin firm working in Web3 tokenization, real-world assets and compliance since 2017, is a consensus-based score that a regulated platform could demand before letting an agent touch anything. It charges for the verdict the way an agent would expect to pay: 0.01 USDC⁵ per evaluation, settled automatically over x402, the emerging standard that lets machines pay for APIs over plain HTTP with no human in the loop. That combination took first place in the existing-projects tier of the Agentic Commerce track at the Algorand Foundation's x402 hackathon⁶ in Berlin on June 6–7 — a 36⁷-hour build sprint worth $3,000 USDC⁸, plus $500 from the Folks Finance bonus track, both listed in the Foundation's official recap and confirmed on BCP's own site.

The event itself drew 373⁹ registrations and 42¹⁰ submitted projects against a $20,000 USDC¹¹ prize pool. And every claim below carries the label the project prints on its own docs: v0.1.0¹², a design exploration, not for production use, not legally certified, with a regulatory mapping no regulator has reviewed and an independent legal review still planned. Statements about partners and compliance alignment are the team's own, not independently verified facts — a caveat that applies to the rest of this article.

A score that gates what an agent may do

Verun starts from a gap it argues no standard currently fills, then answers it with a number. A panel of validators votes on each evaluation and produces a Verun Agent Score from 0 to 1000; the score decides what an agent may do through a regulated chain API, with every evaluation written to Algorand Testnet as a note transaction — a self-payment carrying the verdict in the transaction's memo field, confirmed in seconds at the network's roughly 0.001 ALGO minimum fee and publicly verifiable through the Lora explorer. That anchor is what turns a score into an audit trail: agent identity, score, consensus, operation and timestamp are recorded for replay, which is the design's stated answer to MiFID II Art. 17's record-keeping duty and the EU AI Act's logging and human-oversight requirements.

Sources: 13 algorand.erster.fund, 13 raw.githubusercontent.com

Three operating modes scale the human's role: Discovery has the agent scan validated platforms and send a structured recommendation; Supervised requires a one-click human approval before a single action; Autonomous lets the agent execute freely after approval, with every action still anchored on-chain. The score thresholds themselves (300¹⁸ for reads, 500¹⁹ for transfers and mints, 600²⁰ for orders) remain design choices the team says are exploratory.

Pay-per-verdict, machine to machine

Before the hackathon, Verun gave verdicts away for free — a model the team's own submission notes cannot sustain a network where validators and infrastructure carry real cost. x402 supplies the smallest possible billing surface: a server answers a request with an HTTP 402 and structured payment requirements, the agent signs a payment transaction, retries with an X-PAYMENT header, and a facilitator verifies and settles before the resource is released. Verun's endpoint hands the verification and settlement to GoPlausible²¹, then runs its own validator consensus and on-chain anchor, returning a verdict, anchor and settlement receipt in one round trip — no subscriptions, no API keys, no onboarding. In the project's words, the agent pays and gets accredited.

The live 402 response advertises five schemes, adding an EURQ variant and an EUR→EURD bridge route beyond the three above. On the revenue side, the design routes protocol-side USDC fees into xALGO, Folks Finance²³'s liquid staking token, for native ALGO yield — but that leg is not operational: the treasury endpoint reports a 4.432 ALGO balance with an xALGO position of exactly zero and no opt-in, and the testnet asset ID the treasury module references does not resolve on the testnet indexer. The fee split that would distribute revenue — 70%²⁴ protocol, 10% validators, 10% agent kickback, 10% reserve — is likewise proposed only; no fee contract is deployed.

The trust layer's own trust problem

The credibility of a consensus score rests on who is doing the voting, and there the prototype is thinnest. The current 2²⁵-of-3 validator set is two seats operated by BCP Partners itself (the ERSTER founding validator and an internal test node) plus tokenforge, a genuine German security-token platform — it claims to have issued the first security compliant with Germany's eWpG electronic-securities law in 2022 and lists 50+²⁶ projects across Europe — but one whose integration is explicitly exploratory: in the prototype, the ‘tokenforge’ vote is computed locally from the score rather than by calling tokenforge's API, with a real call deferred to a future mainnet state. The docs reserve a third production slot for a licensed institution — banks, family offices, regulated custodians — with eligibility and bonding mechanics still being designed.

Source: 27 algorand.erster.fund

Until that slot fills, the consensus that gates agents is two of three voices held by the company selling the service. The team appears to be shopping the concept around both chains at once: the same Verun prototype also runs on Stellar testnet at the erster.fund domain, priced at 0.005 USDC per call, and BCP separately announced a strategic collaboration with the Stellar Development Foundation.

What is actually live on-chain

The demo's Step 1 fires a genuine HTTP request and the 402 handshake answers correctly, and the underlying pipeline is demonstrably running. The protocol wallet on Algorand Testnet holds 4.432 ALGO, its health and validator endpoints respond, and its most recent confirmed transactions are the self-payments the anchor mechanism produces — the newest, confirmed August 20, 2026, carries a memo that decodes to a full verun-evaluation payload (agent agt_demo, score 820, consensus LOW, all three validators listed). The prototype has also issued exactly one soulbound credential on testnet: ASA 759213121²⁸²⁹, ‘Verun SBT agt_demo’ (VSBT), a single unit created default-frozen with the issuing address as clawback — matching the minting code in the open-source repository³⁰, which creates a default-frozen Algorand Standard Asset rather than a registry entry.

That last detail exposes a documentation gap worth noting. The published docs describe the SBT as a note-field registry verified through a ‘/data/{key}’ indexer path and show an example issuer address that Algorand's own indexer rejects as invalid — artifacts carried over from the parallel Stellar build. The documented /api/sbt-list and /api/sbt-status endpoints return 404 on the live API, while /api/mint-sbt is implemented. Two other honesty checks: searches of GoPlausible's public discovery registry for ‘erster,’ ‘verun’ and the domain return zero results, so agents cannot currently find the endpoint through the facilitator directory; and the repository has zero stars, no releases, and its last commit is dated June 7 — the hackathon weekend — even as the deployed endpoint still writes anchors. The landing page's investor dashboard, with €4M of a €4M raise shown sold and a 15.2% weekly referral figure, is³¹ illustrative product UI, not live platform data.

What would make it real

The roadmap hangs the next milestones on four changes: a multi-sig, validator-managed treasury; migrating the anchor from note transactions to an ARC-20 smart ASA that carries the verdict tier; mainnet deployment with audited contracts and the pending legal review; and filling the third validator seat with a licensed institution. What is already true is mechanical — the 402 handshake answers, consensus runs, anchors land on-chain, one SBT has actually been issued — but they are testnet mechanics whose commercial and regulatory claims are untested. Verun ultimately faces the same question it poses to every agent it scores: whether an independent party, not the operator itself, vouches for it.