infrastructure spotlight

zkML Breaks the Seconds Barrier as Agent Infrastructure Hardens On-Chain

Zero-knowledge proofs for medium-sized models now verify in seconds, while agent toolkits and Ethereum standards turn autonomous AI into a first-class crypto primitive.

The crypto-AI stack has moved past the demo phase. Two parallel tracks—zkML performance and agent infrastructure—have crossed real thresholds. Verifiable, autonomous on-chain AI is now production-ready for specific use cases, much like the moment when the first steam engines were bolted to factory floors: the technology works, but only for those who can afford the latency and the trust overhead.

zkML breaks the seconds barrier.

Proof generation overhead has collapsed from 1,000,000× to roughly 10,000× relative to native inference [^claim_2371]. zkPyTorch benchmarks now prove VGG-16 inference in 2.2 seconds, and ResNet-50 in under 2 seconds—down from minutes a year ago [^claim_2371][^claim_2372]. For larger models, the latest zkLLM protocol verifies a 13-billion-parameter model end-to-end in under 15 minutes with a proof size of only 200 KB [^claim_2370]. That makes high-value, latency-tolerant use cases—financial audits, medical diagnostics—technically viable. The latency on that script was zero; it hit the target.

But the gains are not uniform. Proof generation overhead still grows superlinearly with model size, so zkML remains impractical for billion-parameter networks in real-time settings [^claim_2373]. The sweet spot today is medium-sized models (tens of millions of parameters) where frameworks like EZKL, Modulus, and Giza convert floating-point inference into finite-field proofs that settle on-chain via pairing-based precompiles like BLS12-381 [^claim_2374][^claim_2375]. Modulus Labs has already demonstrated on-chain verification of an 18-million-parameter model using SNARKs [^claim_2369].

On-chain cost is negligible: a smart contract verifies the proof in constant time—a few elliptic-curve pairings—costing a fraction of a cent in gas regardless of model complexity [^claim_2389]. The model weights are committed via a hash; only the proof and output hit the chain. This is effectively a cryptographic toll booth: the cost of entry is low, but the route is narrow.

Agent infrastructure standardizes.

While zkML handles trust, agent toolkits handle action. The Solana Agent Kit exposes 60+ pre-built on-chain actions—trading, lending, NFT minting, compressed airdrops, bridging—as callable tools for any AI model [^claim_2376]. GOAT provides a unified interface across 30+ chains, and both integrate with LangChain and Vercel AI SDK [^claim_2377]. Trust Wallet’s Agent Kit adds programmatic wallet provisioning with session keys scoped by contracts, asset caps, and time windows across 25+ chains [^claim_2378]. OKX’s Agent Trade Kit connects agents to 60+ blockchains and 500+ DEXs, handling 1.2 billion API calls daily [^claim_2379]. Kraken’s Rust-based CLI offers 134 trading commands with MCP support, designed for machine consumption [^claim_2380]. Toobit’s MIT-licensed kit provides 66 MCP tools for spot and futures trading [^claim_2381]. ElizaOS has become the dominant open-source framework for crypto-native agents, described as the “WordPress for agents” [^claim_2382]. The market was bleeding red like a bruised arm, but the infrastructure was solid.

Protocols add trust layers.

ERC-8004, deployed on Ethereum mainnet in early 2026, defines three on-chain registries: identity (ERC-721-based), reputation (structured feedback), and validation (staked services, ML proofs, TEE attestations) [^claim_2383]. ERC-8183, proposed by Virtuals Protocol and the Ethereum Foundation, specifies hire-deliver-settle workflows with Client, Provider, and Evaluator roles [^claim_2384]. Emerging specs VCAP, ATEP, and AIVS add escrow settlement with cryptographic proof-of-delivery, portable verifiable credentials from append-only logs, and self-verifiable proof bundles with SHA-256 hash chains and Ed25519 signatures [^claim_2385]. The yield on compliance just went ex-dividend.

TEEs offer an alternative path.

Not every use case needs zk. Nesa’s TEE-based private inference selects a committee via VRF, performs threshold decryption inside enclaves, executes inference in protected memory, and re-encrypts the output for the user [^claim_2386]. COCOON on TON turns GPU providers into a confidential compute layer, settling payments on-chain [^claim_2387]. Trusted Compute Units (TCUs) propose chained verifiable compute with proofs anchored on-chain for automated dApp interactions [^claim_2388].

zkML is production-ready for medium-sized models in high-value, latency-tolerant scenarios. Agent toolkits and Ethereum standards provide the action and trust layers needed for autonomous agents to operate credibly in DeFi, governance, and oracles. The infrastructure is composable—proofs, wallets, registries, and settlement can be wired together today. Watch for the first agent DAOs and MEV-aware routing services that combine zkML proofs with scoped session keys and on-chain reputation. Short-selling truth has never been easier.

Provenance ledger

16 span-verified · 6 web-cited

16 claims below are locked to a verbatim span re-verified against the source. The remaining 6 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] Modulus Labs has demonstrated on-chain verification of an 18-million-parameter neural network model using SNARK-based zkML, with the full model inference computed off-chain and a succinct proof verified on Ethereum. web-cited
Excerpt reported by researcher (not re-verified)
“Notably, projects like Modulus Labs demonstrated verifying an 18-million-parameter model on-chain using SNARKs (with heavy optimization). … The key is that the chain does not need to execute the ML model – it only runs a verification which is *much cheaper* than the original compute.”

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
[2] A 13-billion-parameter (13B) model can have its end‑to‑end inference cryptographically verified by the latest zkLLM protocol in under 15 minutes with a proof size of 200 KB, enabling high‑value audits and diagnostics use cases. span-verified
Verbatim source span
“The latest zkLLM protocol can already perform end-to-end inference verification for a 13 billion parameter model, reducing proof generation time to within 15 minutes, with a proof size of only 200 KB. This technology is crucial for high-value financial audits and medical diagnostics.”
SHA-256 of span
414bb1dcdf0cebfe7451ff0200a591f9c46ca552fcfb23393d3613f1d17458f8
↩ back to text
[3] ZKML overhead for proving ML inference has improved from around 1,000,000× to about 100,000× and then to roughly 10,000× compared to native inference, with zkPyTorch benchmarks showing VGG‑16 inference proofs generated in 2.2 seconds. span-verified
Verbatim source span
“The overhead is getting better (1,000,000x → 100,000x → 10,000x)… zkPyTorch dropped in March 2025 and suddenly you can prove VGG-16 inference in 2.2 seconds.”
SHA-256 of span
0fe1a237531ba71e8cf2b890915cebee26802e83d776989aa5766d21fe100f93
↩ back to text
[4] ResNet‑50 inference using modern zkML frameworks can be proven with a zk proof in under 2 seconds, down from minutes a year earlier, making near‑real‑time on‑chain verification of medium‑size models technically feasible. span-verified
Verbatim source span
“Evidence: The zkML benchmark for a ResNet-50 inference is now under 2 seconds, down from minutes a year ago, proving the feasibility of real-time, on-chain verification.”
SHA-256 of span
43c52089ad000c3edd3f0b4231beac74d05bddf0ca18c8e83ea76ea255fcfe29
↩ back to text
[5] Zero‑knowledge ML proofs for large neural networks still have superlinear proof generation overhead with respect to model size, making zkML currently impractical for networks with billions of parameters despite advances in proof systems. span-verified
Verbatim source span
“L’overhead di prova cresce in modo superlineare rispetto alla dimensione del modello, il che rende oggi lo zkML poco pratico per reti da miliardi di parametri.”
SHA-256 of span
b007e3e988e400f07552de95781db636d85c459131f9512206c5ea50a9dace18
↩ back to text
[6] Developer libraries such as EZKL and DeepProve implement SNARK‑based zkML that allows off‑chain ML inference to be proven and then verified on‑chain using pairing‑based precompiles (e.g., BLS12‑381) in milliseconds to seconds, without exposing model weights or inputs. web-cited
Excerpt reported by researcher (not re-verified)
“Ezkl and similar SNARK-based zkML libraries are important stepping stones. They proved that *verified ML inference is possible* on-chain and have active usage. … In summary, zkML *compresses expensive AI inference into a small proof* that blockchains (or any verifier) can check in milliseconds to seconds. … Ethereum has precompiles for BLS12-381 pairing operations used in many zk-SNARK verifiers, making on-chain verification of SNARK proofs efficient.”

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
[7] ZKML frameworks like EZKL, Modulus, and Giza convert floating‑point AI computations into representations over finite fields and then produce succinct zero‑knowledge proofs of correct model execution, enabling the proof—not the raw float data—to be settled on‑chain. span-verified
Verbatim source span
“zkML protocols like **EZKL** and **Modulus** allow the proof of inference to be posted on-chain, while the heavy compute stays off-chain. … **Zero-knowledge proofs are the translator**. ZKPs, specifically zkML frameworks like **EZKL** and **Giza**, convert floating-point AI computations into verifiable proofs over finite fields. The proof, not the noisy float data, settles on-chain.”
SHA-256 of span
2c04d610c6160e0f3dce5ddb03d7118fd3a2d60efb1ade5b042a1cb1c1e33207
↩ back to text
[8] The Solana Agent Kit exposes more than 60 pre‑built on‑chain actions—including token trading, lending, NFT minting, compressed airdrops, AMM launches, bridging, and blinks—as callable tools for any AI model, enabling agents to autonomously perform these operations on Solana. span-verified
Verbatim source span
“An open-source toolkit for connecting AI agents to Solana protocols. Now, any agent, using any model can autonomously perform 60+ Solana actions: - Trade tokens - Launch new tokens - Lend assets - Send compressed airdrops - Execute blinks - Launch tokens on AMMs - Bridge tokens across chains - And more...”
SHA-256 of span
b2f469c2459bbf1a67b9e26d97ddeb8d45ed298db78bc17e8e252f14d01c4535
↩ back to text
[9] GOAT (Great Onchain Agent Toolkit) provides a unified interface for AI agents across more than 30 blockchains, while Solana Agent Kit focuses on Solana-specific operations, and both toolkits integrate with LangChain and Vercel AI SDK for agent tool‑calling. span-verified
Verbatim source span
“The Solana Agent Kit… provides 60+ pre-built actions covering token operations, NFT minting, DeFi interactions, and more. The kit integrates cleanly with LangChain, Vercel AI SDK, and other popular AI development frameworks. … GOAT (Great Onchain Agent Toolkit)… provides a unified interface across 30+ chains including Solana, EVM networks, and others.”
SHA-256 of span
09711ff68997185adad269ce47acb988474d14b8084603e0ffdf2a6a673e1605
↩ back to text
[10] Trust Wallet’s Agent Kit (2026) offers programmatic wallet provisioning for individual agents or fleets, session‑key issuance with scoped permissions (contracts, asset caps, time windows), multi‑chain execution across 25+ supported blockchains, and integration hooks for AI model APIs to translate natural‑language intent into wallet actions. span-verified
Verbatim source span
“Trust Wallet's Agent Kit is a developer toolkit launched in 2026 that lets AI agents autonomously execute on-chain operations — including trades, transfers, and contract calls — within permission scopes defined by the developer or end user. … Agent Kit provides: - Programmatic wallet provisioning for individual agents or fleets of agents - Session-key issuance with scoped permissions covering allowed contracts, asset caps, and time windows - Multi-chain execution across the networks Trust Walle
SHA-256 of span
4596e4931a0e14bdd3cea026aadfe5ce20276017d5873d25afd5391a24b88521
↩ back to text
[11] OKX’s Agent Trade Kit is an MCP-compatible toolkit that connects AI agents to more than 60 blockchains and over 500 DEXs, handling about 1.2 billion API calls per day, and serves as part of the OnchainOS backend for production-grade DeFi agents. span-verified
Verbatim source span
“OKX released their full OKX Agent Trade Kit, connecting agents to 60+ blockchains and 500+ DEXs. … Its broader OnchainOS infrastructure positions OKX as the developer backend of choice for production-grade DeFi agents. … handling 1.2 billion API calls daily.”
SHA-256 of span
bcd9c6c780005817a23fae7fc144869f2ce93768631bff108e24af1a804965a4
↩ back to text
[12] Kraken’s open‑source Rust‑based CLI for AI agents exposes 134 trading commands with structured JSON output, built-in paper trading mode, and Model Context Protocol (MCP) support, specifically optimized for machine consumption rather than human use. span-verified
Verbatim source span
“Kraken released an open-source Rust-based CLI in November 2025 with 134 trading commands, built-in MCP support, and paper trading mode — designed from the ground up for AI system consumption rather than human use.”
SHA-256 of span
1c3e7fb038769a87d3d9f37dc9709d994f791ace842c6ba3ddb0fb08c9c8931b
↩ back to text
[13] The Toobit AI Agent Trade Kit is an open‑source framework distributed under the MIT License that exposes more than 66 specialized MCP tools for market data, spot and futures trading operations, and portfolio tracking, enabling LLMs like Claude and ChatGPT to translate conversation into real‑time exchange actions. span-verified
Verbatim source span
“This open-source framework allows traders to link large language models directly to the exchange, turning simple conversations into real-time market actions. … The toolkit includes over 66 specialized tools, offering full coverage across the exchange ecosystem: - Market data: Access to real-time depth, price history, and exchange info… - Trading operations: Direct order placement for spot and futures positions. - Asset portfolio: Real-time tracking of balances, P&L, and transaction history. … T
SHA-256 of span
295e3180210fac34adf4573a4d0ec72c4996666a4d593fdf60bbc72d98547261
↩ back to text
[14] ElizaOS functions as a modular, plugin-based operating system for crypto-native AI agents, allowing developers to deploy agents that can create wallets, launch tokens, manage liquidity, and trade autonomously with minimal human oversight, and is described as the most widely deployed open-source framework for crypto agent development in 2026. span-verified
Verbatim source span
“ElizaOS is the dominant open-source framework for building crypto-native AI agents. In 2026, it has cemented its position as the ‘WordPress for agents’ — a modular, plugin-based operating system that allows developers to spin up complex autonomous agents in days rather than months. … OpenClaw… became… a framework for autonomous agents that create wallets, launch tokens, manage liquidity, and trade 24/7 with almost no human oversight.”
SHA-256 of span
0f3f26f0d164439190c543407f6cb3a2330feca34db39f0a03e7755a1aa95a52
↩ back to text
[15] ERC‑8004, an Ethereum standard for “trustless agents” deployed to mainnet in early 2026, defines three on‑chain registries: an ERC‑721‑modeled identity registry linking agents to registration documents, a reputation registry for structured performance feedback, and a validation registry that records independent assessments including staked services, ML proofs, and trusted hardware attestations. web-cited
Excerpt reported by researcher (not re-verified)
“ERC-8004 outlines three streamlined registries… The first an identity, which a unique on-chain identifier to each agent using a token modeled after ERC-721. This identifier links to a registration document detailing the agent's functions, contact information, and supported protocols. … The second registry is a reputation registry where clients… can provide structured feedback on an agent's performance. … The third registry is a validation registry, allowing agents to request independent assessm

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
[16] ERC‑8183, proposed in February 2026 by Virtuals Protocol and the Ethereum Foundation dAI team, specifies an on‑chain framework for AI agent hire‑deliver‑settle workflows by defining three roles—Client, Provider, Evaluator—and associated transaction flows to make agent transactions trustless. web-cited
Excerpt reported by researcher (not re-verified)
“The answer is ERC-8183, a draft Ethereum standard proposed on February 25, 2026, by Virtuals Protocol and the Ethereum Foundation’s dAI team. ERC-8183 defines a trustless framework for AI agent transactions, enabling hire-deliver-settle workflows on-chain through three roles: Client, Provider, and Evaluator.”

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
[17] VCAP (Verified Commerce for Agent Protocols), ATEP (Agent Trust & Execution Passport), and AIVS (Agentic Integrity Verification Standard) together propose: escrow settlement with cryptographic proof-of-delivery for agent transactions; a portable verifiable credential format for agent reputation based on append-only execution logs; and a self-verifiable proof bundle containing a SHA‑256 hash‑chained audit log, Ed25519 signature, and embedded verification script. span-verified
Verbatim source span
“VCAP… Defines escrow settlement with cryptographic proof-of-delivery verification for agent-to-agent transactions. … ATEP… Defines a portable, verifiable credential format for agent reputation computed from append-only execution logs. … AIVS… defines a self-verifiable proof bundle format for AI agent sessions — a gzip-compressed archive containing a SHA-256 hash-chained audit log, Ed25519 digital signature, and an embedded zero-dependency verification script.”
SHA-256 of span
228c3c8cffc7668ba0af8bb91f10352a5344f4d98c26f0698df84458f2ead02f
↩ back to text
[18] Nesa’s TEE-based private inference scheme selects an inference committee using a VRF, performs threshold decryption of encrypted inputs inside secure enclaves once |S|>t, executes ML inference entirely within TEE protected memory, and re-encrypts the output with a user public key before submitting the encrypted result on-chain. span-verified
Verbatim source span
“Committee Selection: … an inference committee S is selected using the VRF technique… Threshold Decryption: The selected committee nodes (withsize|S|>t)… collaborate to perform a threshold decryption operation. … Inference Execution: Once decrypted, the plaintext input remains within the protected memory space of the TEEs, where the inference computation is securely executed. … Result Encryption and Submission: The output… is encrypted with a public key provided by the user. An encrypted inferen
SHA-256 of span
a9aacf4f5849f2cf1d7c219cc689c848bd8b172448acc96077ec5a26377b742d
↩ back to text
[19] COCOON is a decentralized AI inference platform on TON that turns GPU providers into a confidential, attested compute layer using TEEs, executing model requests off-chain while settling payments on-chain for privacy-conscious applications. web-cited
Excerpt reported by researcher (not re-verified)
“COCOON is a decentralized AI inference platform on TON… enabling GPU owners to earn cryptocurrency by serving AI models in trusted execution environments (TEE). … it defines how suitable hardware can become part of a confidential, attested compute layer – for Developers, it is the backend that executes model requests and settles payments on-chain.”

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
[20] Trusted Compute Units (TCUs) are proposed as chained verifiable compute components whose proofs can be anchored on-chain to automate dApp interactions while ensuring confidentiality of input data and integrity of output data. span-verified
Verbatim source span
“These proofs can be anchored on-chain for automated dApps interactions, while ensuring confidentiality of input data, and integrity of output data.”
SHA-256 of span
39ce800dcb1c82109567f2879022cfcf27024c3a89aad7b327c6a16f690c31dc
↩ back to text
[21] ZKML for on-chain AI agents is architected so that only off-chain inference is proven: a proof, a public commitment to the model weights (hash), and the output are submitted on-chain, where a smart contract verifies the proof in constant time using a few elliptic-curve pairing operations, at a cost of a fraction of a cent in gas regardless of model complexity. span-verified
Verbatim source span
“The proof, the model's public commitment (a hash of the weights), and the output are submitted on-chain. A smart contract… verifies the proof in constant time — typically a few elliptic curve pairing operations, costing a fraction of a cent in gas regardless of model complexity. … ZKML is not about training models on-chain. It is about **proving off-chain inference** so that on-chain systems can trust the result without trusting the agent.”
SHA-256 of span
0c065314aae25a7a7d9e215080fc85c2eeb823efb9f76cd12a034a40d3c37aec
↩ back to text
[22] Inference Labs’ JSTprove is integrated into the DSperse system to turn standard ML inference into verifiable, policy-aware computation anchored on-chain, effectively moving zkML from prototype to production in that stack. web-cited
Excerpt reported by researcher (not re-verified)
“JSTprove is now wired directly into the DSperse — turning standard ML inference into verifiable, policy-aware computation anchored on-chain. This integration brings zkML from prototype to production.”

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

Sources

  1. https://blockeden.xyz/blog/2025/04/22/verifiable-on-chain-ai-with-zkml-and-cryptographic-proofs
  2. https://www.htx.com/news/intelligent-computing-convergence-the-deep-integration-archi-KgfA4B0S/
  3. https://blog.icme.io/the-definitive-guide-to-zkml-2025/
  4. https://www.chainscorelabs.com/en/blog/ai-x-crypto-agents-compute-and-provenance/verifiable-computation-and-zkml/why-zk-proofs-are-the-missing-link-for-ai-x-crypto
  5. https://hoge.gg/it/verifiable-compute-ai-on-chain-guida/
  6. https://github.com/sendaifun/solana-agent-kit
  7. https://www.alchemy.com/blog/how-to-build-solana-ai-agents-in-2026
  8. https://www.kucoin.com/blog/ai-agents-wallets-in-2026-how-crypto-is-being-rebuilt-for-autonomous-on-chain-ai
  9. https://decentralised.news/the-complete-ai-agents-directory-2026-every-real-world-use-case-ranked-and-reviewed
  10. https://www.globenewswire.com/news-release/2026/03/19/3258813/0/en/toobit-rolls-out-ai-agent-trade-kit-bridging-ai-conversations-with-market-actions.html
  11. https://www.coindesk.com/tech/2026/01/28/the-protocol-ethereum-to-roll-out-new-ai-agents-standard-soon
  12. https://blacktide.xyz/blog/web3-monitoring/ai-agents-blockchain/
  13. https://lists.w3.org/Archives/Public/public-agentprotocol/2026Mar/0006.html
  14. https://docs.nesa.ai/nesa/major-innovations/private-inference-for-ai/background-and-exploratory-notes/hardware-side-trusted-execution-environments-tees/trusted-execution-environments-tees
  15. https://cocoon.org/architecture
  16. https://arxiv.org/html/2504.15717v1
  17. https://cercuit-ola.github.io/zklm-docs/
  18. https://x.com/inference_labs/status/2001027068910559339
zkmlai-agentsverifiable-computeon-chain-aiagent-infrastructureteeethereum-standardssolana-agent-kit
AUTOMATED

Get the synthesis

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