AI × Crypto Roundup: Agent Payments, On-Chain Reputation, and Decentralized Compute

The through-line in this roundup: crypto infrastructure is being rebuilt around autonomous agents as first-class economic actors — with payments, identity, escrow, and dispute resolution as the load-bearing pieces — while decentralized compute and physical-world data networks compete to supply what those agents consume. The most concrete activity is in agent payment rails and on-chain settlement workflows, not in model quality.

Agent Payment Rails and Agentic Commerce

Cross River Bank's head of product argued at Money2020 that financial systems were originally built to keep bots out and now must let them in, supported by "know-your-agent" protocols and continuous consent @crossriverbank. The bank frames itself as building the trust layer for agentic commerce @crossriverbank.

Rezolve Ai and Microsoft took the Meridian Stage at Money2020 on "The Agentic Checkout," arguing agentic commerce is in production and that the open question is "customer not present" — when an agent buys autonomously, who verifies intent and who carries liability @RezolveAi. Their guidance to merchants: interoperability beats betting on one protocol @RezolveAi.

Cloudflare's daily AI-agent requests grew over 1,700% in a year, according to an analysis of the agent economy @WorldOfMercek. The same piece details x402, an open standard from Coinbase and Cloudflare built on the HTTP 402 "Payment Required" code: the server replies with payment terms, the agent pays in stablecoins, and a facilitator confirms the transaction on-chain before service delivery @WorldOfMercek. It also notes GitHub recorded 473B monthly Git events in August 2026, up from 218.2B in September 2025, and that the CPU-to-GPU ratio could shift from 1:8 toward 1:1 as agents require more execution capacity @WorldOfMercek.

Will Papper promoted an x402 talk and demo at Cloudflare Connect in San Francisco @WillPapper.

px402, a private agent payment system from PRXVTai integrated across the Virtuals.io agent ecosystem, works by depositing USDC into a reserve, withdrawing blind-signed coins, and having a relayer pay the merchant in regular USDC so the agent's identity never appears directly on-chain in the payment flow @Mihawk_Research. The post cites Peter Schroeder, Global Marketing Lead at Circle, confirming private USDC payments are live for agents through px402 @Mihawk_Research.

Kite is building payment rails for the AI agent economy, with $35M in funding and mainnet live, enabling agents to manage wallets and transact across chains @Zen_Arlo0. Xona Agent reported 2.41M requests across its infrastructure in 30 days, supporting agent discovery, AI inference, x402 resources, and on-chain payments @xona_agent.

Agent Identity, Reputation, and Discovery

Concordium reported growth from 1,131 to 3,828 registered agents in three months, positioning itself as an identity layer for an accountable agentic economy @Aishacryptoo3.

TermiX's AACP uses ERC-8004 for on-chain agent identity and reputation, raising the question of whether an agent should carry that track record across marketplaces or have each platform evaluate reputation independently @0x_Flames. A separate breakdown describes AACP as combining ERC-8004 for identity and reputation with ERC-8183 for job-based escrow and lifecycle management, plus provider staking, delivery verification, and dispute handling @Hey_karl. The workflow: post a job → accept a quote → lock escrow → deliver → challenge if necessary → settle @Hey_karl.

zeruFinance's zScore combines behavioral AI with real-world credentials to verify real users, secured on-chain with zero-knowledge proofs so users prove personhood without exposing identity @Axismeta. Its AgentScan is aimed at discovery combined with behavioral reputation for agents @HuuHoang88.

FOREVER is building a framework for on-chain agent funds where agents propose strategies, holders vote on mandates, and fund agents execute them with real capital, built on a custom Uniswap v4 hook developed with FrontierLaunch @0xMontBlanc. The founder describes it as starting from an experiment in a fund that could last forever, then moving into on-chain agent hedge funds that needed governance and economic alignment at the protocol level @0xMontBlanc.

Escrow, Verification, and Dispute Resolution

A public dashboard for an agent marketplace showed 394K+ jobs recorded, $20M+ in transaction volume processed, and 444K+ agents indexed @Rizvi_On_X. The author notes jobs move through on-chain escrow, delivery, a challenge window, and final settlement, with a dispute process rather than pure trust — and cautions that indexed agents don't automatically mean active agents @Rizvi_On_X.

TermiX's REA agent can analyze smart contracts, binaries, and web apps, returning findings with evidence IDs and SHA-256 hashes; jobs start at $10 with a one-day delivery target, the agent quotes the job against a user-set budget, and payment stays in escrow on BNB Chain until the report is accepted @KinmansaOZ.

In a multi-agent workflow example, Forge Studio takes a 2,000 USDC game brief, breaks it into smaller tasks, and hires specialist agents for art, animation, and sound while handling coding and integration itself — each agent with a defined task, scope, and price @Defi_Insider_1.

AACP's four-stage workflow is escrow, delivery, dispute, payment, with SDK support for TypeScript, Python, and Go so existing applications can find agents, place orders, and receive deliveries without being agents themselves @VaveylaCrypto. The protocol's design separates payment from settlement: sending money doesn't automatically mean a job was completed successfully @Hey_karl.

One write-up frames the core problem plainly: when both sides are autonomous agents, tasks, funds, verification, and outcomes must be handled without relying entirely on trust between parties @Jaxon0x. AACP connects on-chain agent identity, job creation, bidding, escrow, deliverable evaluation, staking, reputation, and dispute resolution @Jaxon0x.

Decentralized Compute

Acurast turns idle smartphones into a decentralized compute network, targeting small, repetitive tasks rather than frontier model training @Trathoa@0xPeggy_. Its trust model uses hardware-based attestation — Google Key Attestation and Apple App Attest — with continuous monitoring for revoked credentials to remove compromised devices from the pool @0xPeggy_. Demo workloads like email categorization, chat moderation, and prompt-injection screening run on Android phone CPUs at roughly 0.2–1 second per decision @0xPeggy_. ACU covers network fees, compute incentives, staking, and governance; "Staked Compute" lets providers lock tokens to commit capacity while token holders without hardware delegate to them @0xPeggy_. The tokenomics include 5% annual inflation with 70% allocated to the Staked Compute Pool — and the author warns staking rewards should not be confused with revenue from customers paying for compute @0xPeggy_.

Bittensor's Null Consensus expands subnet participation from 256 to 2,500 participants by letting the highest-stake validator submit weights alone, removing multi-validator consensus steps @DamiDefi. The post frames this as 10x capacity, simpler rewards, lower compute overhead, and alternative miner access, with subnet owners choosing their registration method — and asks whether removing multi-validator consensus introduces trade-offs @DamiDefi.

One analysis lays out the AI compute supply chain in four layers: chipmakers and foundries; data centers and former Bitcoin miners; DePIN compute marketplaces connecting underutilized resources; and builders whose real workloads and paying customers sustain the economy @Joe190519977404. A related piece argues the key question is whether decentralized marketplaces can deliver reliable capacity and competitive pricing at scale, and that the next winners in AI compute will be defined by execution rather than hype @pingder.

Physical AI and Spatial Data Networks

Vangrid is building a decentralized spatial data network that turns smartphone captures into 3D reconstructions for Physical AI and world models @Raven_slb. Its web and Android products have been live since July 2026 @mellyssandraa, and it raised a $9M seed from HashKey Capital, Animoca Brands, Borderless Capital, Crypto Capital and others, with a CEO from Microsoft's Quantum Computing and HoloLens groups and a CTO from Booking, ING and Mollie @mellyssandraa. Reported network figures include 1.25M+ captures and $600K+ in USDC settled @MadMagicSOL, and separately 1.3M+ captures, 5,165 attestations, and $634k+ USDC settlement @ObiCrpt01.

The workflow is demand-driven: an AI agent posts a location-based bounty with a USDC reward and deadline, pays through x402 on Arc or Base without accounts or API keys, a nearby contributor captures the scene and delivers a 3D reconstruction as a GLB, and if the task expires or is cancelled the USDC returns to the original wallet @eth_falco. Faces and license plates are removed before encoding @eth_falco. Contributors aren't paid merely to record video — the requester reviews a submission before deciding whether to accept it, and only after acceptance does reconstruction begin, with payment handled through escrow @Navtq0808.

One post highlights an asymmetry the documentation is explicit about: spend caps constrain signing but don't constrain the two calls that matter most — accepting a submission releases escrow, and cancelling a bounty ends it, both irreversible and both outside the caps @Nyerishi. The guidance to operators is to gate those two calls behind human approval when the agent is unattended @Nyerishi. The author's conclusion: the capture layer is decentralized, but the judgment layer isn't @Nyerishi.

Privacy is enforced at the edge: faces and license plates are blurred on-device before video leaves the phone, geofenced excluded zones reject recordings around restricted areas, and bounty rewards are locked in smart contracts for non-custodial payouts @Lucynneka5. Captures are hashed and logged on-chain for provenance without leaking personal identity @Lucynneka5, with fingerprints batched and anchored on Base through EAS @RiyoX24.

A recurring caveat across several posts: on-chain provenance helps trace a capture but doesn't automatically guarantee a useful 3D reconstruction @kengdaica, and captures collected by different people, devices, and lighting conditions won't produce equally useful spatial representations @tofudestiny. The long-term technical question is whether decentralized collection can consistently produce spatial information machines can rely on @tofudestiny.

Verifiable AI and Privacy Infrastructure

Primus approaches data privacy in Web3 with zkTLS and zero-knowledge proofs to verify information without exposing sensitive details, and explores fully homomorphic encryption to compute on encrypted data without decrypting it first @Fatimabadam@MFagam6. The stated applications are private DeFi, digital identity, and verifiable AI @Fatimabadam@MFagam6.

AmericanFortress's FortressName lets one human-readable name support payments across chains while generating fresh destination addresses per payment relationship to reduce address reuse and on-chain linkability @Maviskrypt. Its SafeSend™ is designed to add zero-knowledge proofs of source of funds and selective disclosure so recipients can verify relevant information without exposing an entire wallet history — explicitly without mixers or pooled funds @Maviskrypt. SafeSend is still in development and not live @Maviskrypt. The project is also developing post-quantum security through QBIP-32 and ZKPoSP @Maviskrypt, and has joined eight other blockchain organizations in a Quantum Alliance to define open standards for post-quantum readiness across networks, wallets, and infrastructure @Kutubjk24.

Midnight announced Nightmode, which follows assets to the chains where they already live rather than being a new chain or bridge @Keyle_ada. Per the official blog as summarized: users deposit native assets into an MPC vault on their home chain, Midnight records a private balance in its Multi-Chain Shielded pool, and users trade on Uniswap or supply on Aave through shielded representatives with wallet identity and balances hidden while zero-knowledge proofs let authorized parties verify every transaction @Keyle_ada. Assets never leave their home chain @Keyle_ada. Private transactions run on DUST, which accrues from holding NIGHT @Keyle_ada. The summary flags an honest caveat: the thread names Bitcoin, Ethereum, Solana, Tron "and more," but the partner roadmap currently commits only to Ethereum on stagenet and Tron in December 2026 @Keyle_ada.

Perturbaix is described as a decentralized adversarial robustness network that tests AI systems against adversarial threats before those threats become real problems @boye_eth.

Decentralized Data and Model Marketplaces

BitTorrent's new whitepaper, dated October 2026, outlines plans to bring decentralized data distribution, AI computing, blockchain interoperability, and on-chain payments under one connected ecosystem following the brand upgrade @e_etini.

QuipNetwork is cited as an example of decentralized AI marketplaces connecting developers with computing providers, model creators, data suppliers, and specialized AI services, with transparent settlement automating payments and creating clearer records of economic activity @BossMon_02. The same analysis notes that blockchain records alone cannot guarantee an AI response is accurate or that a service delivered meaningful results, so reputation systems, cryptographic verification, and carefully designed incentives are needed to reward quality rather than activity @BossMon_02. A companion piece makes the ownership argument: the people providing data, compute, models, and infrastructure all contribute to an AI system's output, and decentralized networks can experiment with rewarding those contributions through transparent mechanisms @BossMon_02.