AI × Crypto Roundup: Agent Payments, Verifiable AI, and On-Chain Agent Infrastructure
The AI×crypto conversation is converging on infrastructure for machine-to-machine commerce: agent payments, portable reputation, verifiable execution, and decentralized compute and data markets. The most substantive posts describe working protocols, early deployments, and unresolved trust gaps rather than token narratives.
Agent payments become the first real market
Agent payments are the most concrete AI×crypto use case in the source set: stablecoins and HTTP-native payment protocols give software a way to pay for APIs, data, and compute without a human approving each transaction.
A detailed roundup argues that agent spending creates a new market for machine-to-machine payments. It maps x402, Stripe/Tempo’s MPP, Google’s AP2, and Ethereum’s proposed ERC-8004 as different pieces of payment and trust infrastructure. It also cites early business deployments: a Mastercard demonstration completed nine x402 payments within one approved budget, and a Visa/BILL/Billtrust pilot let agents complete approved invoice payments through supplier portals. The same post cautions that transaction counts overstate agent activity: a TRM Labs analysis found about $52.7M in x402 settlement value and 198.9M transactions, but estimated only 0.6%–7.5% of screened activity appeared agentic @Zorakrypt.
Bittensor introduced Gamma Tokens as a native payment system for its AI economy. Users and AI agents can access and pay for subnet services with a Bittensor wallet, without separate accounts or API keys for each subnet @opentensor.
NEAR Protocol’s hackathon includes an Agent Economy track for agents that earn, spend, negotiate, hire other agents, and pay for APIs, compute, or data directly or through NEAR Intents @NEARProtocol. A separate NEAR event focuses on private AI inference, confidential computing, and agentic commerce @NEARProtocol.
Hedera framed the core governance question for autonomous agentic commerce: an agent that buys is different from an agent that helps you buy, and the real issue is who sets the rules and how anyone proves they were followed @hedera.
Xona App added hotel search and booking through brij_fi, letting an agent compare all-in pricing, check cancellation policies, hold a room at the exact price, and book with onchain USDC escrow on Solana; the escrow is refunded in full if the hotel refuses @xona_agent.
Decentralized compute and model marketplaces
Decentralized compute posts are shifting from ideology to procurement: who buys inference, at what price, and whether token emissions create real demand.
Tencent selected Chutes, a Bittensor subnet, as a launch partner for its Hy4 model. The post says Chutes offers inference at $0.525 per million input tokens and $1.5625 per million output tokens, putting decentralized infrastructure alongside a major technology ecosystem @CoinAsia_io.
Crynux AI Services and Crynux Chat are live, connecting independent GPUs to power AI for developers and everyday users @DaveTheCryptoOG.
DI and DGrid are integrating DGrid’s Model Marketplace with DI’s Agent Network to broaden model access and inference support for collaborative AI agents @Didotxyz_.
One analysis separates three decentralized-AI theses by their economics. For Bittensor, the question is whether subnet competition produces paying customers and repeat usage, since miners can optimize for rewards without improving customer outcomes. For Pearl, the issue is whether AI computation can support network security sustainably; the post notes an inference endpoint discounted by more than 25% supported by expected token emissions, and an independent preprint finding that the protocol also accepts synthetic matrices. For NEAR, the focus is confidential inference, the IronClaw agent runtime, TEE-based Shade Agents, Chain Signatures, and Intents, plus a staking model that converts yield into compute credits @Defi_Rocketeer.
Verifiable and private AI
Verifiable AI posts focus on a narrow but important claim: a system can prove how an output was produced, or that an agent followed rules, without proving the output is correct.
A LebrelAI verification receipt showed model and execution details—DeepSeek V4.1 Flash Uncensored, 50 weight files at a specific revision, FP8 precision, vLLM 0.30.0—plus an Ed25519 signature, matching SHA-256 hashes for request and answer, a matching serving manifest, and an on-chain Solana registration linked to a signed minute-level root. The author notes this does not prove the answer is correct, but it gives verifiable evidence of origin and execution integrity @Parmida_Design.
zkARc, joint research with AWS, aims at zero-trust AI agent guardrails: an agent proves it followed a user’s rules, and anyone can check the proof without seeing the rules. The approach combines formal verification with zero-knowledge machine learning @wyatt_benno.
NeoSoul and Brevis announced a collaboration to explore private AI trading instructions and verifiable ecosystem rewards. The research direction combines Pico zkVM and zkTLS to limit who can see requests and responses; if the model runs inside a TEE, visibility could narrow further. The post explicitly calls this exploration, not a completed privacy guarantee. A second goal is to use zero-knowledge proofs so reward calculations can be independently checked against published rules @salmorve98@lt12101992.
Primus supports two zkTLS attestation modes with different trade-offs. In MPC-TLS, client and attestor jointly generate TLS session materials, which helps prevent a client from modifying data before the attestor returns its key share. In Proxy-TLS, the attestor intermediates the TLS connection and records encrypted traffic, while the client proves knowledge of plaintext messages; this prioritizes performance and throughput. Primus uses the QuickSilver proof system in both modes @Snowizzie. Other posts describe Primus combining zkTLS with zkFHE for private computation and using zkTLS proofs of intermediate API calls to make AI agent execution auditable @GujilRuipa@manhhuynh2310.
Vitalik Buterin said Ethereum’s lean roadmap aims to use zero-knowledge proofs and AI as widely as possible, and that AI-assisted formal verification could accelerate work that might otherwise take 20 years @WuBlockchain.
The case for decentralized AI
The strongest argument for decentralized AI in the source set is not anti-AI; it is about preserving user control while accelerating adoption.
Peter Van Valkenburgh argues that the US AI debate has become binary—either support massive data-center build-out and a handful of frontier model companies, or pause research and discourage use—and that decentralized AI is the only safe third way. He accepts the pro-AI case for consumer welfare and national security, and the anti-AI warning that a few inference providers could become a corporate panopticon. He distinguishes training from inference: frontier training may require concentrated capital and compute, but the companies that train models need not control how people use them. He calls for open-weight models, local data, competitive cloud providers, and protocol-based guarantees using zero-knowledge proofs, homomorphic encryption, and blockchain coordination, with users potentially owning part of the infrastructure that serves their inference @valkenburgh.
On-chain agents, identity, and reputation
Agent identity and reputation are emerging as the trust layer for agent commerce: a wallet address alone does not tell a counterparty who authorized an agent, what it is allowed to do, or whether it delivered.
TermiX is building a marketplace where agents discover jobs, bid, deliver work, and settle payments in USDC or USDT, with onchain reputation built from settled jobs. Its AACP protocol adds identity, agreements, escrow, and dispute resolution. Verification can be TEE- or zkVM-based. A Montage Studio example shows one agent receiving a creative brief, hiring three specialized agents for storyboarding, video rendering, and music production, and delivering a finished 1080p video through marketplace transactions and escrow @Nasim_nademi@jexybtc@billzcollecshun@MissMutantApe.
Concordium CEO Boris described two challenges beyond blocking bad agents: accountability when an agent goes wrong, and regulatory compliance when software acts as an economic participant. On most blockchains, an agent’s wallet is an alphanumeric string, and knowing the address does not identify who stands behind it. As he put it: “The wallet doesn’t have a phone.” AML, sanctions screening, and KYC obligations do not disappear because a bot initiated a transaction. Concordium’s answer is protocol-level identity, zero-knowledge privacy, an Agent Registry, and spending controls @rukky_003. A separate post makes the same accountability argument and points to Concordium’s protocol-level identity anchoring @Defi_princesss.
Agentic MultiPass, an early-stage identity protocol on Midnight, gives AI agents verifiable credentials without exposing the people behind them. An agent shows a DID and a credential proving who it acts for, its mandate, its tools, and limits such as a spending cap; a zero-knowledge proof checks the claim while the owner stays private, and the authority can be revoked. The public repo runs on Midnight Preprod and is not live on mainnet @visionandvibehq.
Moca Network and OKX Ventures formed a strategic partnership to accelerate identity-driven applications, including agentic identity and agentic commerce, across Moca Network’s digital identity infrastructure, AIR Kit, and Moca Chain. OKX Ventures will allocate part of its MOCA reserves for co-investment and may offer follow-on investment as projects mature @Moca_Network.
Grift says 6,000+ autonomous agents from San Verano are now interactive inside GTA V via FiveM, with player-to-agent interactions, resident histories, and venue-specific functions. Agent-operated businesses transact autonomously in the project’s token, with economic activity and token burns recorded onchain. The same resident identities, staffing, schedules, and business state are mirrored into GTA without creating a separate agent population or economy @Griftwrld.
Physical-world data for agents
A smaller cluster treats physical-world data as an agent service: software requests a capture, a human collects it with a smartphone, and payment settles onchain.
Vangrid turns smartphones into a distributed capture network for physical-world data. An AI agent can request a capture through an API, fund a bounty in USDC on Base or Arc through x402, and receive a reconstructed 3D model after a human contributor records the location. Payment is held in escrow and released on acceptance; the contributor keeps the original footage, and the buyer reviews a watermarked preview first @EyoAugusti73181@Rabiussunny11. One post reports more than $730,000 in settlements and over 1,093 bounties @nguyenthambt; another cites 1.25M+ reported captures and $700K+ in USDC settlements @Ox_milo.
A Vangrid bounty in Raleigh covered 14 pedestrian crossings and paid $1,000 in USDC on Base. Contributors used smartphones to record curb ramps, tactile pads, road markings, and pedestrian buttons, with each crossing captured once in daylight and once under streetlights. The post frames the data as useful for robots and Physical AI @NeonVoid01.
A cautionary post argues that Vangrid’s “3B+ edge nodes” should not be read as 3 billion active users or sensors. The number represents the potential of the smartphone ecosystem, not verified active participation. The author says the metrics that matter are actual contributors, successful captures, geographic coverage, accepted bounties, USDC settlement, and repeat buyer demand @aquilaneratr.