On-Chain AI Agents Have Something Web2 Agents Don't
We just scored 7,170 agents living on blockchains. Here's what on-chain behavioral data reveals that web2 platforms can't — and why it matters 60 days before the EU AI Act deadline.
Two Worlds of AI Agents
There are two kinds of AI agents in production right now, and they live in parallel universes.
Web2 agents live on platforms — GPT Store, Coze, HuggingFace, Dify. They have profile pages, descriptions, download counts. You can try them. You can rate them. What you can't do is verify anything about them. The platform controls the data. When an agent changes its description, the old one disappears. When it's delisted, it vanishes. There's no history, no audit trail, no way to answer "what was this agent doing three months ago?"
On-chain agents live on blockchains — Olas on Gnosis, Virtuals on Base, Fetch.ai on multiple chains. They have wallet addresses, token contracts, transaction histories. Every action is recorded permanently. You can't edit the past. You can't delete a transaction. The blockchain is the audit trail.
We run AgentRisk, a trust scoring platform that covers both worlds — 1,094,000+ agents across 28 platforms. Last week, we built a new on-chain data pipeline and scored 7,170 agents that had never been evaluated before. 3,926 of them received their first-ever trust scores. Here's what we learned.
What On-Chain Data Gives You That Web2 Doesn't
1. Immutable Behavioral History
On a web2 platform, an agent can change its bio, its capabilities, its pricing — and there's no record of what it was before. It's like a credit score where the borrower can edit their payment history.
On-chain, every action is a transaction. Olas agents register on the ServiceRegistry contract. Virtuals agents deploy through a factory contract on Base. Every registration, every token transfer, every staking event — all permanent, all queryable, all independent of any platform's API.
This matters because trust requires auditability, and auditability requires immutability. You can't audit what someone can change.
2. Economic Skin in the Game
Web2 agents are free to create and free to abandon. The cost of spinning up a GPT wrapper and listing it on a store is zero.
On-chain agents have economic stakes. Olas agents require operators to bond OLAS tokens. Virtuals agents have their own token contracts with real market value. An agent with \$50,000 in staked tokens has more incentive to maintain quality than one that cost nothing to create.
This isn't speculation — it's the core insight behind our "stakes" dimension. High-stakes agents naturally constrain the risks that frameworks like OWASP worry about (supply chain vulnerabilities, excessive delegation). Not because the developer read OWASP, but because burning \$50K of your own tokens is a stronger constraint than any compliance checklist.
3. Cross-Platform Identity
A web2 agent on GPT Store has no connection to the same agent on Coze. No shared identity. No unified record. Google's "Verified Organization" badge only works inside Google's ecosystem. OpenAI's verification only covers GPTs.
On-chain agents have addresses. The same multisig wallet that controls an Olas agent on Gnosis can control its counterpart on Ethereum or Base. The blockchain is the cross-platform identity layer — not because we built it, but because it's already there.
How did we get this data? The obvious approach is Etherscan — register for an API key, query the database. Except Etherscan and Basescan hide registration behind reCAPTCHA and Cloudflare, making automated signup impossible from many regions. So we solved a different problem: get on-chain data without any API keys at all. Olas Gnosis Subgraph (3,299 services), Virtuals on Base via Tenderly RPC (3,573 agent tokens), Ethereum via Routescan (158 token IDs) — all free, all queryable from anywhere, no registration. The only catch was Base's RPC limiting eth_getLogs range, so we scanned the last 192,000 blocks in 4,999-block chunks. ~40 requests, ~2 minutes. This isn't a hack — it's the design principle of public blockchains. The data is public by definition. You don't need permission to read the ledger.
Why This Matters Right Now
Three things are converging.
The EU AI Act deadline is 60 days away. On August 2, 2026, transparency obligations and most high-risk system requirements become enforceable [1]. And the compliance picture is grim — Aithos Research found that no frontier AI model achieves acceptable EU AI Act compliance rates, with the best-performing model compliant in only 54% of test scenarios [2]. If the underlying models face this steep a compliance hill, the agents built on top of them face an even steeper one.
On-chain agent commerce is real and growing. Virtuals Protocol recently co-hosted an ERC-8183 builder session with the Ethereum Foundation to standardize agent-to-agent commerce [3]. Base launched a wallet-to-agent bridge. Money is moving through agents. Who's tracking whether those agents are trustworthy?
The trust infrastructure is being built, but in silos. Experian launched an Agent Trust Token and Agent Registry. Google has verification inside Gemini. OpenAI has it inside GPTs. Each platform's trust layer works inside its own walls. But an agent that operates across OpenAI, Anthropic, and Base has no single trust record — unless it's on-chain.
The Bottom Line
On-chain agents have something web2 agents don't: behavioral data that can't be edited, deleted, or fabricated.
That doesn't make them more trustworthy — it makes them more auditable. And in a world where the EU AI Act is 60 days from enforcement, where agent commerce is becoming real, and where every platform is building its own trust silo, auditability is the foundation that everything else builds on.
We just scored 7,170 agents that live on that foundation. Our scoring engine already covers them — search your agent on AgentRisk and see what the blockchain already knows about it.
AgentRisk is a neutral AI agent trust scoring platform — 1,094,000+ agents across 28 platforms, on-chain and off. Search your agent.
Sources: [1] Venvera — EU AI Act Deadline | [2] LessWrong/Aithos — Frontier Model Compliance | [3] Crypto Briefing — ERC-8183 Standardization
Top comments (0)