I read fourteen writeups of MoonPay's newest integration for AI agents handling crypto lending on Solana over the past two weeks — trade-press summaries, exchange blog posts, four Substack takes, a handful of Twitter threads presented as analysis. They all told me the same three things, and none of them told me the fourth. Solana sits at a $43.46B market cap and a $74.56 spot price against an all-time high of $259 set on 2024-12-18, and that context matters for what these pieces do not say. What none of the coverage addresses is where the private key actually lives at the moment an autonomous agent moves collateral on your behalf. That is the piece I want to write.
Two weeks of reading and I still cannot tell you, from any of those fourteen articles, whether the agent holds a signing key, borrows a session-scoped delegation, or fires transactions through a custodial API that MoonPay signs. That is not a small omission. That is the entire question.
What They All Get Wrong
The shared analytical error is treating "AI agent" as if it were the product. It is not. The product is a delegation model — a specific answer to the question of who signs a transaction and under what constraints — and the "AI agent" is the thing consuming that delegation. Every piece I read spent the first two-thirds of its word count explaining what an autonomous agent is, why Solana's throughput makes it a good venue for programmatic lending, and why MoonPay's onramp reach makes this "a big deal for retail." None of that is wrong. It is also not the story.
Here is the pattern. The writeup opens with the announcement. It quotes the press release. It gestures at Solana's speed — sub-second finality, sub-cent transaction fees, the usual talking points. It notes that Solana's market cap sits somewhere near $43B and that SOL trades around $74 against a $259 ATH from December 2024, and it treats those numbers as color rather than context. Then it lists three lending protocols the agent "can interact with" and closes with a sentence about how this "unlocks new possibilities for retail participation in DeFi." That is the template. Fourteen articles, one template.
The error inside the template is subtle and it is the same everywhere. Every one of these pieces conflates two very different things: an agent that *recommends* a lending action to a user who then signs, and an agent that *executes* a lending action under a pre-signed authorization. Those are not adjacent products. They are different risk models entirely. The first is a chatbot with API access to a price oracle. The second is a programmatic delegate holding, or borrowing, spend authority over your assets. The trade-press coverage collapses both into "AI agent handles lending" and moves on, because the distinction is annoying to explain and interrupts the momentum of the piece.
The second error, downstream of the first: none of the coverage separates the *onramp* layer (MoonPay's actual specialty — moving fiat into crypto) from the *lending execution* layer (a completely different discipline involving collateral management, liquidation risk, and oracle exposure). MoonPay's competence at the first does not translate to the second. That is not a criticism of MoonPay. It is a category question, and the coverage does not ask it.
The third error is the beginner-audience framing. I understand why it happens — the top exchanges recommended for beginner personas (Coinbase, Kraken, Bybit, Bitget, OKX in that order by persona rationale) exist because first-time crypto users need hand-holding, and any AI-agent product marketed to that audience gets covered as if it were another hand-holding tool. But lending is not onboarding. A beginner who signs a delegation they do not understand and then wakes up to a partial liquidation because SOL dropped 8% overnight is not being onboarded. They are being processed.
What Is Almost Always Missing
The custody boundary. Every piece I read described "the agent" as if it were a discrete entity floating somewhere between the user and the blockchain. It is not. There is always a key, and that key lives somewhere specific, and the somewhere-specific determines the entire risk profile of the product. This is the single most important operational fact about any autonomous on-chain agent, and it is the one fact the coverage never states.
The signing model has, at most, four plausible shapes for a product of this kind. One: the user retains the key and the agent proposes transactions that the user signs — this is a recommendation engine wearing an agent's hat, safe but not really "autonomous." Two: the agent operates a session-scoped delegated authority — Solana's account model supports this pattern, and it means the user has signed a bounded permission ("this program can spend up to N SOL from this account until timestamp T") that the agent consumes. Three: the agent's execution runs through a custodial API where the platform holds the underlying key and signs on behalf of the user based on the agent's decisions — this is technically MoonPay-signs, not user-signs, and it changes the custody picture completely. Four: some hybrid where a hardware co-signer at the user's end has to approve categories of transactions.
Not one of the fourteen pieces I read told me which of these four the integration uses. That is the article. Everything else is preamble.
The second missing piece: the liquidation path. Solana-based lending protocols use on-chain oracles for collateral pricing, and those oracles have documented failure modes — stale prices during network congestion, price impact from thin secondary markets, cascade risk when a single collateral asset gets liquidated across multiple positions simultaneously. When an autonomous agent is managing collateral on behalf of a retail user, who decides the liquidation-avoidance threshold? Is the agent authorized to add collateral without asking? Is it authorized to unwind positions early to preserve principal? What happens if the agent's inference layer is degraded — model outage, API rate limit, RPC endpoint failure — during a fast market? None of these questions appear in any of the fourteen writeups.
The third missing piece: the qualified-custodian question. In the U.S., custody providers registered under the NY DFS Trust Company framework — Coinbase Custody, Fidelity Digital Assets — operate under specific fiduciary rules. Anchorage Digital, holding the OCC Federal Trust Charter as the first federally-chartered crypto bank, operates under others. An autonomous agent that moves assets on your behalf sits somewhere in the fuzzy space between "custody" and "software you use." The regulatory posture on that space in 2026 is not settled, and the coverage does not mention that it is not settled.
The fourth missing piece — and I acknowledge I have not seen the technical documentation that would let me answer this cleanly — is what happens on revocation. If a user wants to fire the agent, what is the actual operation? Is it a signed revocation transaction on-chain? Is it a support ticket? Is there a lag? Every autonomous system needs a documented off-switch and the coverage does not ask what the off-switch looks like.
What I Would Say Instead
The frame that would actually help a reader think about this product is not "AI agents are coming to Solana lending." The frame is: *what have you delegated, to whom, under what constraints, and how do you take it back*. That frame works for any autonomous on-chain product regardless of whether the autonomy comes from an LLM, a rules engine, or a keeper bot written in 2021. The AI-agent framing is a rhetorical wrapper; the delegation model is the actual thing.
Concede the strongest point of the enthusiast case up front: yes, an autonomous agent that continuously monitors your lending positions and adjusts collateral in response to volatility is, in principle, better than a retail user who checks their position twice a day and misses the liquidation event that happens at 3am local time. That is a real value proposition. Solana's throughput and finality — the reason it keeps coming up in these writeups — genuinely does make it a plausible venue for programmatic collateral management in ways that congested L1s are not. Fine. I will grant all of that.
But the value proposition depends entirely on the answer to the custody question, and the answer to the custody question determines whether this product is closer to Coinbase Custody's institutional prime-brokerage stack (assets held by a qualified custodian, decisions made by a delegated party under a governed mandate) or closer to giving a stranger your seed phrase and hoping they act in good faith. Those are not adjacent points on a spectrum. They are different products sold under the same marketing.
Here is what I would want a reader to walk away with. First, before you touch any AI-agent lending product on Solana or anywhere else, read the docs section on transaction signing. If the docs do not have a transaction-signing section, that is the answer — do not use the product. Second, look for the word "delegation" and look for what scopes and expirations the delegation supports. A delegation without a time limit and without a spend cap is not a delegation, it is a surrender. Third, look for the revocation path. If the revocation path is a support form, that is a custodial product marketed as non-custodial. Fourth, look at the liquidation policy — specifically, whether the agent is authorized to add collateral from your onramp balance without a fresh authorization. If yes, understand that your MoonPay fiat balance is now downstream of an oracle you do not control.
I want to be honest about what I do not know. I have not been able to pull MoonPay's technical integration spec for the Solana lending agent product — the announcement material I found is marketing, and marketing does not answer signing-model questions. I am not going to fabricate that specification and I am not going to pretend the trade-press summaries are a substitute for it. What I can tell you is that the shape of the coverage — fourteen articles, one template, zero mentions of the signing boundary — is itself a signal. If the custody model were straightforward and defensible, the press release would say so, and the writeups would quote it. The silence in the coverage is either laziness on the writers' part or ambiguity on the product's part, and I do not yet know which. The honest answer is I need the technical docs, and until I have them, the framing I have offered here is the most useful thing I can give you: a checklist of what to look for and why the absence of any item on that checklist should end your interest in the product.
The open question I keep coming back to is whether the market is going to reward products that answer these delegation questions clearly, or whether the enthusiasm for "AI agents on Solana" will let the ambiguity ride for another cycle. If you have seen a version of this product from any vendor that documents the signing boundary in plain language — not marketing language, actual API-reference language showing scopes and expirations and revocation semantics — send it to me. I have not found one yet, and I would like to be wrong about that.
FAQ
Where does the private key actually live when an AI agent moves my collateral on Solana?
That depends on the specific integration and the honest answer is that the mainstream coverage does not tell you. Plausible models range from a session-scoped delegation you sign once (the agent then consumes a bounded permission), to a fully custodial arrangement where the platform holds the key and signs on the agent's behalf, to a hybrid with a hardware co-signer at your end. Before using any such product, find the docs section on transaction signing. If it does not exist, treat that as the answer.
Is MoonPay a qualified custodian for the assets an agent manages?
MoonPay is best known as a fiat onramp provider, not as a chartered custodian in the sense that Coinbase Custody (NY DFS Trust Company), Fidelity Digital Assets (NY DFS Trust), or Anchorage Digital (OCC Federal Trust Charter) are. Whether an AI-agent lending product built on top of a MoonPay integration inherits any custodial protections depends on the specific legal wrapper of the product, which the announcement coverage does not specify. Read the terms of service, not the press release.
What happens if the agent's inference layer goes down during a market move?
This is one of the questions no coverage I read addresses. A responsibly-designed autonomous system needs a documented failover — what the agent does when its model, its RPC endpoint, or its oracle feed is unavailable during volatility. Options include unwinding positions to cash, holding collateral steady and pausing new actions, or handing control back to the user with a notification. If the product docs do not specify the failover behavior, assume the worst case and size your exposure accordingly.
Does Solana's throughput actually make it safer for programmatic lending?
Sub-second finality and sub-cent transaction costs do make Solana a plausible venue for continuous collateral management in ways that congested chains are not — an agent can rebalance positions cheaply and quickly. But throughput does not solve the oracle problem, the liquidation-cascade problem, or the delegation-model problem. Those are protocol-design and product-design questions that sit on top of the base layer. Fast rails do not fix a bad signing model.
How do I revoke an agent's authority if I want to fire it?
Look for this in the docs before you sign anything. On Solana, a properly-designed delegation is revocable via an on-chain transaction the user signs — instant, verifiable, no counterparty required. A revocation path that runs through a support ticket, a customer service email, or a "please contact us" flow is a custodial product wearing a non-custodial marketing wrapper. Instant on-chain revocation is the standard to demand; anything less is a downgrade.
Is this suitable for someone new to crypto?
The persona data our coverage of beginner-friendly exchanges points to — Coinbase, Kraken, Bybit and similar — reflects a specific set of onboarding needs: simple UX, clear education, low minimums. AI-agent lending on Solana is a different discipline. Autonomous collateral management under oracle-price liquidation risk is not a beginner product no matter how clean the interface is. If you do not yet understand what a health factor is on a lending protocol, do not delegate that decision to software.
What is the difference between an agent that recommends and one that executes?
An agent that recommends produces a suggestion which you sign — you retain the key, you retain the decision, and the risk of the agent being wrong is bounded by your willingness to sign. An agent that executes holds a signing authority, delegated or custodial, and moves assets without a fresh human decision at each step. These are not adjacent products. Coverage that treats them as the same category is doing the reader a disservice. Verify which one the specific product is before using it.
Why does Solana's price context matter here?
Because collateral management under leverage is a function of volatility, and SOL's price history — trading around $74 against a December 2024 all-time high of $259 — implies the kind of drawdown that liquidates positions. An autonomous agent handling SOL-collateralized borrowing is operating on an asset that has demonstrated the ability to lose most of its value inside a single cycle. The agent's liquidation-avoidance policy is not a footnote in this context; it is the product.