Most AI agents on a blockchain today are analysts and co-pilots. They read data, suggest moves and run narrow, well-defined jobs, such as rebalancing a yield position or paying for a service in cents. One crypto VC report estimates automated software is behind about 19% of on-chain activity, but most of that is simple bots, and full autonomy is still rare.
What are AI agents genuinely doing on-chain right now?
Agents today do three things with real evidence behind them: they read and summarise on-chain data, they manage simple DeFi positions, and they make small automated payments. DeFi, or decentralised finance, means lending, trading and savings services run by smart contracts, which are programs that live on a blockchain. Anything that needs open-ended judgement, such as discretionary trading, is where the results get thin.
The table sorts the common claims by how much evidence sits behind them, as of November 2026.
| Task | Where it stands | What the evidence shows |
|---|---|---|
| Reading and summarising on-chain data | Common, but error-prone | In an academic benchmark, the best tool-using model answered 67.4% of crypto research tasks correctly, against 80% for junior human analysts |
| Managing yield and liquidity positions | Real, small, early | A crypto VC report counts more than $39 million in liquidity positions run by agents, with most deployments still in testing |
| Paying for services per request | Real, tiny amounts | A crypto investment firm counted $73 million across 176 million agent transactions over twelve months |
| Open-ended trading | Mostly hype | In one human-versus-agent trading contest, the top humans beat the top agents by more than five times |
| Fully autonomous operation, end to end | Not yet reached | The VC report says most agents still work at analyst or co-pilot level |
The pattern is clear. Agents do well where the task has a clear target and clear rules. They struggle where the task is ambiguous, adversarial or needs judgement about what information to trust.
Why does the “one in five” figure overstate what agents do?
The roughly 19% figure counts automation of every kind, and most of it is narrow bot activity rather than reasoning agents. A researcher at the firm behind the report said most of that share comes from simple jobs, such as profiting from the order of trades in a block and routing stablecoin transfers. Truly agentic activity, in that account, is a minority.
The same report found that bots generate more than 76% of stablecoin transfer volume. A stablecoin is a token designed to hold a steady value, usually one US dollar. Moving stablecoins between addresses on a schedule is useful work, but it is not an AI deciding anything.
That distinction matters. A bot follows fixed rules someone wrote. An agent is given a goal and decides the steps itself. Counting both under one label makes "AI on-chain" sound far more advanced than it is.
Headline counts of agents have the same problem. A June 2026 academic study of ERC-8004, an Ethereum standard for registering agent identities on-chain, looked at Ethereum and two other large networks. Only 3%, 4% and 15% of registrations respectively had a valid registration file and a declared service. Most were placeholders.
What does “autonomy” actually look like for an on-chain agent?
Autonomy is a scale, not a switch, and most agents sit near the bottom of it. A useful way to read any product claim is to place it on one of three levels:
- Analyst. The agent reads prices, wallet balances and contract data, then reports or recommends. A human makes every decision and approves every transaction.
- Co-pilot. The agent prepares the transaction and a human approves it. Many products sold as "AI trading" work this way.
- Autonomous. The agent decides and signs on its own, inside whatever limits its owner set. This exists, but mostly for narrow jobs such as rebalancing one yield position or paying per request for data.
The crypto VC report placed most agents at the first two levels and said true end-to-end autonomy has not been reached. The payment data points the same way. Agents are making very many very small payments: $73 million across 176 million transactions works out to well under a dollar each. That is machines paying machines for small services, not agents moving large sums on their own judgement.
Where autonomy does exist, it tends to sit inside tight limits. A yield agent can be allowed to move funds between a short list of approved lending pools and nowhere else. That is real autonomy, but it is closer to a well-fenced robot than a fund manager.
Where does the hype outrun the evidence?
The biggest gap between claim and evidence is trading. Claims that an agent will consistently beat the market have little support in public data. In the human-versus-agent contest cited in the VC report, every account started with $10,000, and the best humans outperformed the best agents by more than five times.
A separate model-versus-model contest, also cited in the report, found that only the top three models had a positive expected return. Agents that held positions for two to three hours did better than ones that traded often. That suggests discipline and fewer trades, not more intelligence, explained much of the gap.
Yield is the stronger case. The report found that one yield agent beat standard lending rates, even after fees. The report adds that it has not been shown to work at larger scale. Rule-based optimisation suits agents. Open-ended strategy, so far, does not.
Reading data is the quiet weak spot. In a 2025 academic benchmark of crypto AI agents, the best models without tools managed at most 28% accuracy. With tools they reached 67.4%, still below the 80% scored by entry-level human analysts. The authors found models often trusted blogs and social posts over the chain itself.
What could go wrong?
The main risk is mistaking a co-pilot for an autopilot, or an autopilot for a reliable one. Labels mean little. A product called "autonomous" may be a simple script. A product called an "assistant" may hold permission to sign transactions without asking you.
On a blockchain, most transactions cannot be reversed. An agent that misreads data, which the benchmark above shows is common, can act on a wrong number with no undo button. A short track record makes this worse. The VC report notes most agent deployments are still in testing, so a few good months say little about how an agent behaves under stress.
There are also risks that grow with adoption. The report flags strategy convergence, where many agents reach identical positions from the same data. It also flags that copying a public strategy can erode its returns, and that opaque agents carry insider risk, because their creators could misuse them. None of this is a reason to panic. It is a reason to ask what the agent actually does.
The trust takeaway
Judge an agent by what it is allowed to do, not by what it is called. Today the honest description of most on-chain agents is "a capable assistant with narrow powers". The ones worth trusting are clear about which level they operate at, show their work, and have limits you can see.
Before you rely on any agent on-chain, check:
- Which level is it really? Analyst, co-pilot or autonomous, in plain words from the provider.
- Can it sign without you? If so, for which actions and up to what amount.
- How long is the track record? Months of results on real funds, not a backtest or a demo.
- Can you see every decision? A full log of what it read, what it decided and what it sent.
- Where does its data come from? Whether it checks figures on-chain or trusts whatever it finds online.