The AI x crypto pitch is moving too quickly from “agents need tokens” to “agents should control money.”
That skips the hard part.
CoinTelegraph’s source context cites veteran macro investor Jordi Visser arguing that AI agents need “food,” and that the food is tokens. The idea is directionally useful: if AI agents become active software workers, they may need to buy data, pay for compute, access APIs, settle microtransactions, interact with tokenized assets, or move value between services.
But the first serious bottleneck is not whether agents can hold crypto.
They can.
The harder question is whether an AI system can understand what kind of crypto it is holding, what risks come with it, where it can move, and what permissions govern the transaction.
That is where the current market is still underbuilt. CoinGecko’s work on rehypothecated-token categories and API changes shows that even human-facing crypto data still needs cleaner labels. Ethereum’s L1/L2 roadmap shows that network routing is becoming more layered. CoinGecko’s tokenomics tools point to another basic need: structured information about supply, unlocks, and distribution.
For AI commerce to work, crypto cannot just be programmable.
It has to be legible to machines.
Agents Need More Than a Balance
A human can look at a wallet balance and ask questions.
What token is this? Is it native? Is it wrapped? What chain is it on? Is there enough liquidity? Is this a scam asset? Can I use it for payment? Is it tied to a DeFi position? Does it depend on a bridge? What contract am I approving?
Humans do not always ask those questions, which is how many losses happen. But at least they can.
AI agents need that context in structured form before they act.
A token balance by itself is not enough. An agent making payments or managing access needs to know whether an asset is appropriate for the task. A stable-value payment token is different from a volatile utility token. A native asset is different from a wrapped claim. A bridged token is different from the original asset. A rehypothecated token may depend on another layer of collateral or protocol risk.
If the system cannot tell the difference, it cannot safely automate decisions.
This matters for small businesses as much as developers. Imagine a business using an AI agent to pay vendors, buy software credits, or route customer refunds. The agent should not be free to choose any asset that appears available in a wallet. It should be constrained by payment policy: approved tokens, approved chains, daily limits, counterparties, fee thresholds, and required human review for anything outside the normal pattern.
Without that, “AI payments” becomes a nicer phrase for automated treasury risk.
Asset Labels Become Operating Instructions
CoinGecko’s planned changes around rehypothecated tokens are not only relevant to DeFi investors.
They are relevant to AI systems.
CoinGecko said DeFi’s evolution requires updated methodology for tracking and ranking assets, including wrapped and rehypothecated tokens. For human users, clearer labels help reduce confusion. For AI agents, those labels may become operating instructions.
A machine-readable label could tell an agent that an asset is native, wrapped, bridged, staked, rehypothecated, or otherwise dependent on another structure. A policy engine could then decide what is allowed.
For example:
- Native ETH may be approved for gas and settlement. - A bridged token may be blocked from treasury payments. - A rehypothecated asset may be excluded from collateral. - A token with poor classification data may require human review. - A volatile asset may be allowed for trading but not vendor payments. - A tokenized claim may require a specific redemption path before use.
That is the kind of infrastructure AI commerce needs.
Not because labels eliminate risk. They do not. But labels let software apply rules before funds move.
Right now, too much crypto context is trapped in blog posts, dashboards, token pages, explorer labels, and user assumptions. AI agents need that context normalized through APIs, wallets, and application permissions.
Otherwise, they will act on incomplete information at machine speed.
A bad idea, now with uptime.
Tokenomics Data Matters for Automated Decisions
CoinGecko’s May update highlighted tokenomics tools that help users understand distribution and unlocks.
That may sound like investor research, but it also matters for AI-connected financial systems.
If an AI agent is selecting assets for payment, treasury routing, collateral, or access, tokenomics can affect whether the asset is suitable. Supply concentration, upcoming unlocks, incentive-heavy usage, and unclear distribution can all change risk.
A human analyst may decide those details are irrelevant for a small transaction. That is fine. The point is that the decision should be explicit.
An agent should not treat every token with liquidity as equally usable. It should be able to apply different rules depending on the use case.
A token used for a small API payment may only need basic liquidity and sender approval. A token used for working-capital reserves needs a much higher standard. A token used as collateral in an onchain market needs valuation, liquidation, and dependency checks. A token selected for recurring vendor settlement should probably avoid unclear supply dynamics and thin markets.
That requires data.
More specifically, it requires data that software can consume cleanly.
The AI x crypto market often talks about autonomous agents as if the biggest challenge is intent. But for financial use, the bigger challenge may be context. A system can know what it is trying to accomplish and still pick the wrong rail if the asset data is weak.
Layered Networks Need Clear Payment Routes
Ethereum’s L1/L2 roadmap also matters because AI agents will not be operating in a simple single-chain world.
Ethereum.org has described the goal of scaling Ethereum as a cohesive system, with L1 and L2s using their different strengths to create a stronger platform. For human users, that means better usability is still a work in progress. For AI agents, it means route selection becomes a core safety issue.
If an agent needs to make a payment, where should it settle?
On L1? On an L2? Through a bridge? In a stablecoin? In ETH? Through a tokenized instrument? Through a custodial platform? Through a smart-contract wallet with spending rules?
Each route has different fees, timing, liquidity, finality assumptions, and operational risks.
A human may tolerate some friction while choosing a network. An agent needs a policy framework. It needs to know which networks are approved, which assets are accepted, where the recipient can actually receive funds, and whether moving between layers creates risk or delay.
This is especially important for small businesses.
A business does not want an AI agent saving a few cents on fees by routing payments through a network the accounting system cannot track or the recipient cannot use. It does not want funds stranded on the wrong chain. It does not want automated bridge decisions hidden inside a “pay” button.
For AI commerce, chain abstraction cannot mean risk abstraction.
The user may not need to see every technical step, but the system must still enforce the right rules.
Wallets Need Agent-Specific Permissions
A normal crypto wallet is built around direct human approval.
AI agent wallets need something more granular.
The right model is not “agent has wallet” or “agent has no wallet.” It is scoped authority. The agent should have permission to do certain things, with certain assets, on certain networks, within certain limits, for certain counterparties.
That could include:
- Spend up to a daily limit. - Pay only approved vendors. - Use only approved tokens. - Operate only on approved networks. - Require human review for new contracts. - Block unlimited token approvals. - Reject unknown or poorly labeled assets. - Log every transaction with a reason. - Pause activity when risk data is missing.
This is not exotic. It is standard operational control translated into crypto.
The problem is that many current wallet experiences still treat permissions too broadly. For AI use, broad permissions are dangerous. An agent should not be able to approve open-ended token spending because a contract asked nicely. It should not bridge assets without policy approval. It should not interact with unknown DeFi contracts just because a yield looks attractive.
Agent wallets should be more like corporate spending cards than personal hot wallets.
Limited, monitored, and revocable.
Data Providers Become Part of the AI Stack
Crypto data providers are usually discussed as research tools.
In an AI-agent market, they become infrastructure.
Agents will need prices, liquidity, tokenomics, asset classifications, contract labels, network status, fee estimates, and risk metadata. They may need to distinguish real tokens from spam tokens, native assets from derivatives, approved contracts from unknown ones, and normal payment routes from high-risk paths.
That makes API quality more important.
If bad data reaches a human dashboard, it may mislead a user. If bad data reaches an automated wallet, it may trigger action. The consequences are different.
CoinGecko’s evolution from a simple price tracker into a broader crypto companion, along with its tokenomics and asset-classification work, reflects the direction of the market. Data is no longer just for watching prices. It is becoming part of how products decide what users hold and what software should be allowed to do.
AI will raise that standard again.
The better the data layer, the safer the agent layer can become. The weaker the data layer, the more every AI wallet becomes a pile of assumptions with a signing key attached.
What Readers Should Watch
Watch whether AI x crypto projects explain their permission model. If the answer is vague, the product is not ready for real funds.
Watch whether wallets add spending limits, asset allowlists, contract restrictions, and transaction logs designed for agents.
Watch whether token data becomes more machine-readable. Labels around wrapped, bridged, staked, and rehypothecated assets matter.
Watch L1/L2 routing. AI payment systems need clear settlement paths, not just cheaper transactions.
Watch tokenomics data. Automated systems should not treat supply structure as irrelevant.
Watch human override controls. Serious AI commerce should make it easy to pause, review, and revoke authority.
The Grounded Takeaway
AI agents may eventually become meaningful users of crypto rails.
But the real opportunity is not simply giving agents tokens. It is building payment and data infrastructure that lets software act within clear limits.
That requires machine-readable asset risk, better token classifications, reliable network routes, tokenomics context, wallet permissions, and human oversight. Without those pieces, AI commerce will be faster than normal crypto payments but not safer or smarter.
The technology story is not that agents need wallets.
It is that wallets, APIs, and networks need to become disciplined enough for agents to use them without turning automation into a loss multiplier.
That is less flashy than the AI-token pitch.
It is also the part that actually matters.