infrastructure spotlight

Verifiable AI Agents Hit Mainnet: The Trust Premium Just Collapsed.

Zero-knowledge proofs of 13.5B-parameter models now cost $0.001 and complete in 15 seconds, while toolkits give agents gasless access to 200+ on-chain actions across EVM and Solana—the infrastructure for autonomous DeFi is shipping.

In the digital antiquity of late 2024, a thing happened that future historians of trustless computation will note with the same reverence we reserve for the first steam engine: Lagrange Labs’ DeepProve-1 cranked out a zero-knowledge proof for a complete inference run of OpenAI’s GPT-2, all 13.5 billion parameters, in a mere 15 seconds, spending gas fees so tiny ($0.001) they barely register on the balance sheet of a protocol[^3198][^3205]. This is not a parlor trick with a shrunken model; it’s the first production-hardened zkML system that eats entire large language models for breakfast, spitting out proofs verifiable across Ethereum, Solana, and Sui via a Rust SDK[^3198]. The age of verifiable inference as a DeFi primitive has begun.

The proof generation pipeline itself is a ruthless assembly line. JSTprove, built atop Polyhedra’s Expander Compiler Collection, grabs a standard ONNX model by the throat and forces it into a ZK circuit, handling quantization and proof generation with the clinical efficiency of an assassin—no private parameters exposed, no secrets leaked[^3199]. For those who need answers now, before the next block lands, PHAZE applies cryptographic hashing like a dead drop, then smacks on a ZK proof to certify early exits from fat baseline models, delivering a compact certificate of correct execution with the low latency of a sniper’s trigger[^3207]. These tools aren’t academic; they mean that an AI’s outputs can be proven on-chain in seconds, not the glacial hours of yesteryear[^3198][^3199][^3207].

But the raw proof is only valuable if it can be traded on. Enter the agent toolkits, the market makers of this new synthetic economy. GOAT (Great Onchain Agent Toolkit) is a bazaar of over 200 on-chain actions: Uniswap swaps, Jupiter routes, OpenSea mints, Polymarket bets—spread across EVM, Solana, Aptos, and the rest, plugging into LangChain, Eliza, ZerePy like a trader’s Bloomberg terminal[^3202]. It turns an AI model into a native capital allocator, able to deploy assets or analyze data with the cold precision of a quant fund. And 0xGasless AgentKit layers on account abstraction and gasless transactions, so agents can transfer tokens, read balances, and soon deploy ERC-20s and mint NFTs without end-users ever touching a gas token[^3203]. The agent becomes invisible infrastructure, the plumbing through which capital flows, and you never even see the fees.

Beneath this, protocol layers are rising like the trading floors of a new financial district. NEAR is staking its claim as the headquarters for the agent economy, hawking cross-chain unified liquidity, confidential execution, private inference, and a throughput north of 1 million TPS—with its private inference already humming in apps like Abound and Brave[^3200]. Ritual operates as a decentralized clearing house: independent Checkers validate AI results, Gateways connect smart contracts to models, building a trust-minimized orchestration that would make a 19th-century clearinghouse blush[^3201]. And Supra’s Threshold AI Oracles push further, embedding AI reasoning directly on-chain, fusing inference and execution into a single workflow that delivers cryptographically verifiable answers to smart contracts like a secured payload[^3204]. This is the settlement layer for automated truth.

The composability yields a stack where every piece is marked to market. A developer can deploy a Solidity executor on a ZK rollup, pull the strings of off-chain AI via ethers.js, sip data through Covalent API, route intents with Tria SDK, and inject live ML outputs from Chainlink oracles, all for transaction costs that round to zero on a hedge fund’s ledger[^3210]. The AI’s decisions are auditable via DeepProve or JSTprove, and its actions executed gaslessly via 0xGasless across chains via GOAT, turning the agent into a self-auditing, fee-less trading engine[^3198][^3199][^3203][^3202].

The trust premium that centralized AI extractors have been collecting is about to evaporate. Market-making agents can now publish a cryptographic receipt proving they ran a specific model on specific inputs, shooting information asymmetry in the head like a double-crossing spy[^3198][^3207]. Governance votes can be triggered by verified predictions, not just vibes[^3204]. MEV-aware strategies can publish their inference audits for all to see, turning black-box trading into an open book[^3198]. And with NEAR and Ritual serving up private inference, confidential strategies can finally operate on-chain without exposing their edge[^3200][^3201]. The informational arbitrage that DeFi has been running on is being closed.

The stack for verifiable autonomous agents is no longer a pitch deck fantasy. With proof costs at a $0.001 rounding error and tooling covering over 200 actions, the race is on to spin up the first fully on-chain AI fund that publicly attests to the reasoning behind every single trade[^3198][^3202]. Watch for DeepProve’s 2025 roadmap as it extends these proofs to LLaMA and Claude, bringing trust guarantees to model training and fairness—and thus pricing the value of AI honesty into every contract[^3206]. The yield on verifiable intelligence just went ex-dividend. Get in or get out.

Provenance ledger

12 span-verified · 2 web-cited

12 claims below are locked to a verbatim span re-verified against the source. The remaining 2 are web citations: the URL was checked, but the excerpt is the researcher's summary and was not re-derived from the page. Citation markers in the text jump here.

[1] The PoIm-based on-chain learning framework for DeFi attack mitigation performs gas-prohibitive training computation on Layer-2, propagates verified model updates to Layer-1, and enables gas-bounded, low-latency inference inside Ethereum smart contracts, with quantized and loop-unrolled implementations of logistic regression, SVMs, MLPs, CNNs, gated RNNs, and formally verified decision trees that are bit-exact to off-chain models and proven in Z3. span-verified
Verbatim source span
“We present the first decentralized, fully on-chain learning framework that: (i) performs gas-prohibitive computation on Layer-2 to reduce cost, (ii) propagates verified model updates to Layer-1, and (iii) enables gas-bounded, low-latency inference inside smart contracts. … We develop quantization and loop-unrolling techniques that enable inference for logistic regression, SVM, MLPs, CNNs, and gated RNNs (with support for formally verified decision tree inference) within the Ethereum block gas l
SHA-256 of span
a58f43b88299583cc2e3a9c808901a1a3c84dbb82054dd99dd7d05a7c61f1363
↩ back to text
[2] DeepProve-1 generates a zero-knowledge proof for the full inference of OpenAI’s GPT-2 using Groth16, producing a proof over 13.5 billion parameters in 15 seconds with gas fees of $0.001, and supports multi-chain verification on Ethereum, Solana, and Sui via a Rust SDK, with total value locked across Lagrange-served projects reaching $100 million by 2025. web-cited
Excerpt reported by researcher (not re-verified)
“DeepProve-1 successfully verifies OpenAI's complete inference of GPT-2, becoming the first production-ready zkML system. … generating a proof with 13.5 billion parameters in 15 seconds, with gas fees of only $0.001, far lower than the $10-50 cost of traditional cloud computing. … The system utilizes Lagrange's Rust SDK in conjunction with the Groth16 proof protocol, supporting chains such as Ethereum, Solana, and Sui. … By 2025, Lagrange serves over 50 projects, with a total value locked (TVL)

This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.

↩ back to text
[3] JSTprove provides an end-to-end zkML pipeline that accepts ONNX models, quantizes them, compiles them to circuits via Polyhedra Network’s Expander Compiler Collection, and automatically generates and verifies zero-knowledge proofs of AI inference, turning conventional inference into transparent, auditable computation without exposing model parameters. span-verified
Verbatim source span
“The JSTprove public launch introduces a full end-to-end pipeline for zero-knowledge AI proofs. It accepts ONNX models and automates proof generation and verification through Polyhedra’s Expander Compiler Collection. … The process involves quantizing an ONNX model, generating circuits via Expander, and creating proofs that confirm the model executed as intended without exposing private parameters.”[5]
SHA-256 of span
acd028000828ac6badc683ced5ef7ad4a1dd894aa8ae49249a9acb470cef65db
↩ back to text
[4] NEAR’s agent infrastructure claims to unify liquidity across more than 35 chains, provide confidential execution and private inference, and scale to over 1,000,000 TPS while offering a secure agent harness for deploying AI agents with private inference services used by applications including Abound, Brave, and Venice AI. span-verified
Verbatim source span
“NEAR is the open infrastructure powering the agent economy. NEAR unifies liquidity across 35+ chains, keeps execution and inference confidential, and scales to over 1 million TPS. … NEAR’s open infrastructure combines cross-chain execution, confidential settlement, private inference, and a secure agent harness. … Deploy agents using the private inference infrastructure serving Abound, Brave, and Venice AI.”[9]
SHA-256 of span
4435f343ad72cae0111edae23dd55933774c8da5955603c089287b0b43235fa8
↩ back to text
[5] Ritual’s decentralized execution layer for on-chain AI introduces a permissionless network where independent actors serve, call, and validate models, with distinct roles including Checkers to verify results or model configurations and Gateways that connect smart contracts and off-chain clients, enabling trust-minimized AI inference orchestrated on-chain. web-cited
Excerpt reported by researcher (not re-verified)
“Ritual is a decentralized execution layer for running and verifying AI inference. It introduces a permissionless network where models can be served, called, and validated by independent actors just like nodes, relayers, or oracles. … Checkers verify results or model configurations. … Gateways connect smart contracts and offchain clients. … Just trust-minimized AI inference, orchestrated onchain.”[11]

This excerpt was not re-derived from the source page, and may paraphrase or condense it. Check the source before relying on it.

↩ back to text
[6] GOAT (Great Onchain Agent Toolkit) exposes more than 200 on-chain tools across EVM chains (Base, Polygon, Mode, Sei, etc.), Solana, Aptos, Chromia, Fuel, Sui, Starknet, Zilliqa and others, and integrates with agent frameworks like AI SDK, LangChain, Eliza, ZerePy, GAME, and ElevenLabs to let AI agents perform DeFi, minting, betting, and analytics operations directly from agent workflows. span-verified
Verbatim source span
“GOAT (Great Onchain Agent Toolkit) is a library that adds more than +200 onchain tools to your AI agent. - +200 tools: DeFi (Uniswap, Jupiter, KIM, Orca, etc.), minting (OpenSea, MagicEden, etc.), betting (Polymarket, etc.), analytics (CoinGecko, BirdEye, Allora, etc.) and more. - Chains: EVM (Base, Polygon, Mode, Sei, etc.), Solana, Aptos, Chromia, Fuel, Sui, Starknet, Zilliqa, etc. … - Agent Frameworks: AI SDK, Langchain, Eliza, ZerePy, GAME, ElevenLabs, etc. … The open-source framework for c
SHA-256 of span
dda69f5351aaf9c271036f64de8cc7097a00d01efda2c9fe534b693e900bc61f
↩ back to text
[7] 0xGasless AgentKit enables AI agents to interact with EVM chains using account abstraction and gasless transactions, supporting on-chain actions such as reading wallet details and balances, transferring and trading tokens, with planned support for deploying ERC-20 and ERC-721 contracts, minting NFTs, performing DEX swaps of ERC-20 coins, and wrapping ETH to WETH on Base. span-verified
Verbatim source span
“0xGasless AgentKit lets you build AI agents that can trade, manage funds, and interact with DeFi protocols in natural language. … It enables gasless transactions and account abstraction on EVM chains. … Supported Onchain Actions: - Getting wallet details and balances - Transferring and trading tokens - Coming Soon: - Deploying ERC-20 tokens - Deploying ERC-721 tokens and minting NFTs - Buying and selling Dex Swap ERC-20 coins - Wrapping ETH to WETH on Base. … AgentKit is a toolkit that gives AI
SHA-256 of span
28719c21d02335091f59ad3a99e4e9a164de54347e7f6d7ce62611d354fd1ef5
↩ back to text
[8] Supra’s Threshold AI Oracles embed AI reasoning natively on-chain, fusing inference and execution into a single decentralized workflow that delivers cryptographically verifiable AI answers to smart contracts, removing centralized middlemen and enabling dApps to consume AI-derived insights in real time on Supra’s vertically integrated stack. span-verified
Verbatim source span
“Threshold AI Oracles… brings a wide spectrum of intelligence onchain natively for your smart contracts on Supra, so you can build dApps that think, decide, and adapt to the real world in real-time. … It embeds AI reasoning natively on-chain with cryptographically verifiable answers, all without centralized middlemen, messy workarounds, or compromises. … Threshold AI Oracles eliminate that friction by fusing inference and execution into one seamless, decentralized workflow that delivers AI insig
SHA-256 of span
59c51582272f15e9d76ff7e72a95659fc04a7cc609d7dcde9b52b7733d68642c
↩ back to text
[9] DeepProve-1 is described as the first production-ready zkML system to successfully generate a cryptographic proof of a full large language model inference, specifically a full inference of OpenAI’s GPT-2, establishing feasibility of end-to-end LLM verification via zk proofs. span-verified
Verbatim source span
“Lagrange Labs is proud to announce DeepProve-1, the first production-ready zkML system to successfully generate a cryptographic proof of a full large language model (LLM) inference. … With DeepProve-1, we’ve successfully generated a zero-knowledge proof for the full inference of OpenAI’s GPT-2…”[8]
SHA-256 of span
1b81448f8c17a073e42ac86de7d0be9a862b8d19eccd5020cbe7d6b06e211fe7
↩ back to text
[10] The DeepProve roadmap for 2025 targets support for major LLMs such as LLaMA and Claude, private and confidential inferences, and new proof types that can cover model training, fine-tuning, fairness, and reasoning, extending zkML from pure inference verification to broader lifecycle guarantees. span-verified
Verbatim source span
“@lagrangedev has released its 2025 roadmap for its zkML framework - DeepProve - focusing on scaling verifiable AI… Key goals include support for major LLMs (like LLaMA and Claude), private and confidential inferences, and new proof types - covering model training, fine-tuning, fairness, and reasoning.”[14]
SHA-256 of span
c655d280ca566d33e216e4dd5739681b1f08180bf15434aea1059ed6518c099e
↩ back to text
[11] The PHAZE framework combines cryptographic hashing and zero-knowledge machine learning to generate succinct zero-knowledge proofs of correct execution of ML inference computations C(x,w)=y, where x is public input, w is private witness (model parameters and activations), and y is public output, enabling certifiable early-exit mechanisms from large baseline models while preserving low-latency triggers. span-verified
Verbatim source span
“We propose PHAZE… a novel framework built on cryptographic techniques like hashing and zero-knowledge machine learning (zkML) to achieve low latency inference, via a certifiable, early-exit mechanism from an arbitrarily large baseline model. … An application of ZKPs for generating a succinct proof, \pi, of the correct execution of an ML model’s inference computation, C(x,w)=y, where x \in L is a public input… w is the private witness… and y is the public model output.”[4]
SHA-256 of span
eef54e4380935aae0fc8c02c7d7596f19ccc7b77c4ed36ecc3a20152329308ce
↩ back to text
[12] Cortex (CTXC) embeds machine-learning capabilities directly in smart contracts to pioneer on-chain AI inference, and zero-knowledge machine learning (ZKML) enables an AI provider to submit both the output and a cryptographic proof that the output was generated by a specific model on a given input, enabling on-chain verification of off-chain computation. span-verified
Verbatim source span
“Cortex (CTXC)… pioneered on-chain AI inference by embedding machine learning capabilities within smart contracts… A major step toward trustless decentralized AI is enabling on-chain verification of off-chain computation. Zero-knowledge proofs for machine learning (ZKML) allow an AI provider to submit not only an output but also a cryptographic proof that the output was generated by a specific model on a given input.”[6]
SHA-256 of span
4cf1636f09e61556b04d06b2d1bd43cc7c4163106a9ada16fc1ff7329d939c14
↩ back to text
[13] SynFutures’ on-chain AI framework for DeFi introduces Synthia, an AI trading agent that executes natural-language swap and transfer commands, sources best prices across multiple protocols, intelligently splits orders, and will be extended via a modular SDK and no-code tools for deploying custom agents, culminating in a meta-agent that can automatically create, orchestrate, and manage multiple agents for complex workflows. span-verified
Verbatim source span
“The framework… its first product—Synthia, an AI trading agent—is now live. … AI agents that enable users to swap and transfer crypto assets using natural language commands (e.g., ‘swap 100 USDC for ETH’)… streamlines order execution, sources the best prices across multiple protocols, intelligently splits orders… Phase 2: An open framework that allows… developers to deploy custom AI agents… with no code or a powerful SDK… Phase 3: The introduction of a ‘meta agent’ capable of automatically creat
SHA-256 of span
fca8e6d67739009f324f8899d73168bbbeeb82d8f540da22230fe1d1d6a5da18
↩ back to text
[14] A practical Web3 AI-agent dev stack combines Solidity executors, Viem/ethers.js for off-chain orchestration, data feeds via Covalent API, intent-based routing via Tria SDK, deployment with Hardhat on Base, and recommends ZK rollups to achieve sub-$0.01 transaction costs, with Chainlink oracles used to inject live ML outputs on-chain. span-verified
Verbatim source span
“- Solidity for the executor - Viem/ethers.js for off-chain orchestration - Covalent API for feeds - Tria SDK for intents… - Hardhat: npx hardhat run scripts/deploy.js --network base - Gas tip: Use ZK rollups for sub-$0.01 txns. - Edge: Add Chainlink oracles for live ML outputs.”[10]
SHA-256 of span
09fc553c7e8508acc225419eda1dcd4742a5820785ee9b3e173d2e09eefa8745
↩ back to text

Sources

  1. https://arxiv.org/abs/2510.16024
  2. https://www.binance.com/en/square/post/28856397934873
  3. https://web.ourcryptotalk.com/news/jstprove-public-launch-brings-verifiable-ai-to-developers
  4. https://near.org/
  5. https://medium.com/@angell_art/ritual-the-missing-execution-layer-for-onchain-ai-3014d34827a1
  6. https://github.com/debridge-finance/goat
  7. https://github.com/0xgasless/agentkit
  8. https://supra.com/academy/threshold-ai-oracles-brings-intelligence-on-chain/
  9. https://lagrange.dev/blog/deepprove-1
  10. https://www.hozk.io/news/zk-ai-latest-2025-06-25
  11. https://ar5iv.labs.arxiv.org/html/2511.12592
  12. https://arxiv.org/html/2505.07828v1
  13. https://chainwire.org/2025/02/20/synfutures-unveils-comprehensive-ai-framework-for-defi-launches-ai-trading-agent-synthia/
  14. https://dev.to/estelleatthenook/the-rise-of-ai-agents-in-web3-a-quick-devs-guide-to-on-chain-autonomy-181k
verifiable-computezkmlai-agentsdefiaccount-abstractioninfrastructurenearritualsupra
AUTOMATED

Get the synthesis

AI×crypto research, repackaged with every claim hash-locked to its source. New arXiv → analysis in ~3 hours.