I’ve noticed that a lot of talk is going around about AI agents on the blockchain right now, but few people dig into why they actually don’t work the way we’d like. Galaxy Research has released an interesting report, and it lays bare the core of the problem.

It all starts with a paradox: blockchain is, by definition, programmable, permissionless, and an open ecosystem—so it would seem to be an ideal environment for autonomous agents. But the trouble is that blockchain was created for consensus and deterministic execution, not for machines to understand the economic meaning of what’s happening.

That’s the key difference from traditional algorithmic systems. A typical algorithm can scan the DeFi market, find new contracts, allocate capital—automatically. But the moment an unfamiliar interface appears, the system stalls. It needs a human who can work through the code, understand the mechanics, and write the integration. The human interprets; the algorithm executes.

Agents based on large language models are shifting this boundary. They can read unfamiliar code, analyze documentation, and infer the system’s economic functions. It sounds powerful, but there’s a catch: they do it imperfectly, and in an environment involving real assets, a mistake can cost money.

Galaxy identified four main sources of friction. The first is discovery. On-chain, all deployed contracts look the same to the protocol, but the agent has to figure out which are legitimate, which are fakes, and which are abandoned projects. Humans solve this through websites, social signals, and interfaces. The agent only sees raw bytecode.

The second friction is verification. Remember the story of WETH. On Ethereum, there are almost 200 tokens named “Wrapped Ether,” with the symbol “WETH” and 18 digits after the decimal point. How will a machine determine which one is real? There’s no built-in concept on the blockchain of “this is the official application contract.” Repositories and trusted interfaces help humans, but for an agent, it turns into a logical puzzle.

The third friction is data. Imagine you want to compare yield between Aave v3 and Compound v3. Both are lending markets, and the economic concepts are the same. But the ways to obtain the data are completely different. In Aave, you first need to get a list of reserves, and then, for each one, make a separate call for liquidity and rates. In Compound, each deployment is its own market, and there’s no unified reserves structure at all. The agent has to use different methods for each protocol. This isn’t just inconvenient—it creates delays and introduces the risk of data desynchronization.

The fourth friction is execution. When you click a button in an interface, you’re informally checking: does this look reasonable? What’s the risk? Is slippage acceptable? People do this intuitively. Agents have to encode these checks explicitly. They have to translate the goal “maximize yield with risk control” into a concrete plan: choose the protocol, the market, the volume, and the sequence of actions. Then verify that every step complies with the constraints. And finally, make sure that the result actually matches the goal, even if the transaction is technically successful.

These aren’t just engineering problems—they’re structural contradictions. Blockchain is designed to guarantee the correctness of state transitions, but it does not guarantee that economic states are easy to interpret, that contracts are standardized, or that goals can be achieved.

Part of the problems are a consequence of openness and the lack of permission (which is both a strength and a weakness of blockchain). Part comes from the current state of tools and infrastructure. But the main point is that all existing infrastructure assumes that there’s a human in between interpreting the state and executing an action.

Galaxy suggests that the solution will require new layers: unifying the economic state across protocols, indexing services for semantic primitives, registries for verifying contracts and tokens, and execution frameworks with hard-coded constraints.

As agents begin to truly manage capital, the architectural assumptions of the current interaction level will become increasingly obvious. It’ll be interesting to see how the blockchain ecosystem adapts to this. It seems there will be a lot of building ahead.