Financial Management is Moving From Human Hands to AI Agents
The Agent Economy, Piece 2: The $1.8 Trillion Blind Spot
There is a number that keeps showing up in pitch decks and research reports: $30 trillion. That is Gartner’s projection for the total economic activity mediated by AI “machine customers” by 2030.[1] Gartner forecasts $201.9 billion in agentic AI spending this year alone, a 141% increase over 2025.[2] Goldman Sachs projects that agents will account for more than 60% of the total software economy by the end of the decade.[3] If you have spent any time in crypto, the pattern is familiar: enormous numbers used to justify enormous bets before anyone pauses to check whether the small numbers add up.
So I checked the small numbers. The total amount of capital currently managed by autonomous AI agents in decentralized finance is somewhere around $200 million. That is an educated guess: nobody tracks this as a single category, and the figure shifts week to week, but I mapped every platform I could find with confirmed on-chain activity, from Giza to Theoriq to Almanak, added up what is sitting in vaults and flowing through agent transactions, and that is where the number lands. For context, it is about the size of a small community bank’s asset base, roughly 0.2% of total DeFi TVL. The agent economy, as it exists today, is a rounding error inside a niche.
I believe this rounding error will matter more than anything else happening in finance. Not because the current numbers are impressive, but because what agents represent, software that manages money autonomously, with no sleep, no emotion, and no operational ceiling, is where financial services ends up. But conviction without evidence is just a pitch deck. If you want to understand where autonomous money management is actually headed, you need an honest picture of where it stands right now, and that picture is more interesting than any projection can capture.
This is the second piece in a series I am calling The Agent Economy. The first examined what happens when AI agents get hacked, a question that becomes urgent once real capital is on the line. This piece is the landscape map: which protocols are live, what they are actually doing with money, what infrastructure is being built around them, and one very large gap involving the biggest asset in crypto.
The Protocols That Are Live
I spent weeks mapping every DeFi agent platform with confirmed on-chain activity. Not announced. Not in testnet. Not launching soon. Actually live, managing real capital, executing real transactions. The list is shorter than the conference panels suggest, but the protocols on it are doing genuinely interesting work.
Giza is the clearest proof point. Their ARMA agents on Base handle autonomous yield optimization for stablecoins, and the platform has grown from roughly $300,000 in managed capital to over $32 million in under a year.[4] DefiLlama tracks their TVL at approximately $21 million, though Giza’s self-reported Assets Under Agent figure runs higher.[5] The model is straightforward: agents monitor lending rates across protocols like Aave, Morpho, and Compound, then rotate deposits to wherever the spread exceeds the cost of moving. That growth trajectory, from five figures to eight in twelve months, is steeper than early Morpho vaults over the same period. Re7 Capital, a digital asset investment firm managing approximately $800 million across strategies, recently adopted Giza’s agents for institutional treasury management, making it the first institutional fund to deploy autonomous agents for portfolio allocation.[6]
Theoriq launched AlphaVault in November 2025 and attracted $50 million in TVL within its first week.[7] Their architecture is worth paying attention to because it directly addresses the reasoning-layer vulnerability I wrote about in the first piece of this series: they use “policy cages” that bound what an agent can decide, not just what it can execute. That distinction matters. Whether the TVL held through the market downturn of early 2026 is unclear, as updated figures have not been published, but the architectural approach puts them in a small group of teams thinking seriously about agent-level risk.
Olas runs a fundamentally different model. Rather than managing pooled capital in vaults, Olas coordinates networks of specialized agents that transact directly with each other, and they have surpassed 10 million cumulative agent-to-agent transactions.[8] That makes them the first network to prove that crypto-native micropayments between autonomous agents work at scale. The value is not sitting in a vault. It is flowing through coordination. Think of it as the difference between a bank and a marketplace: Olas is building the marketplace.
HeyElsa has quietly become one of the most active agent platforms on Base, processing over 18.9 million prompts and facilitating more than $503 million in total transaction volume.[9] It functions more as an AI co-pilot than a fully autonomous agent: users direct it rather than letting it operate independently. But the volume is real and on-chain, and it suggests something important about where the market actually is right now. The first useful agents may not be fully autonomous. They may be translators between what you want and how DeFi works.
And then there is Almanak, who launched their token in December 2025 and TVL hit $120 million. The protocol had over 100,000 users, leveraging eighteen specialized agents. Real technology.[10] By mid-February 2026 though, TVL had fallen below $2 million, and the token had dropped 98.7% from its peak. Anyone who lived through 2021-2022 DeFi recognizes the pattern: the capital was tethered to token incentives, and when the token price collapsed, the economic rationale for staying collapsed with it. The agents kept running. The capital they were optimizing evaporated beneath them. Almanak’s technology is real, and the team is still building. But the episode underscores a challenge the entire sector faces: a meaningful portion of agent-managed capital exists because of token subsidies, not because autonomous management generates enough organic return to justify the capital on its own merits. That gap between subsidized and organic returns is the central tension of DeFAI right now.
What You Can Do With an Agent Managing Your Money
Strip away the projections and the jargon, and what DeFi agents do today is remarkably concrete. The best way to explain it is to show you.
Imagine you have savings sitting in a checking account earning close to nothing. In traditional finance, the best you can do is move it to a high-yield savings account, maybe earning 4-5%, and check back in a year. In DeFi, there are dozens of lending platforms, each offering different interest rates that fluctuate based on supply and demand, sometimes changing every few minutes. The best rate right now might be on Aave. In an hour, it might be on Morpho. Tomorrow, it could be somewhere else entirely. Capturing the best yield means monitoring all of these platforms continuously, calculating whether the improvement justifies the cost of moving your capital, executing the transaction, and doing it again the next day. A person could technically do all of this, but not efficiently, not consistently, and not at 3am when the best rate shifts to a protocol they have never heard of. An AI agent can do it around the clock without breaking a sweat.
I put this to practice myself. I connected my wallet to Giza, chose Base as my chain (rates were running up to 15% APR), and then hit the screen that made the whole thing click for me: Auto or Custom. Auto means you hand the agent the keys. It picks the lending markets, sets the parameters, diversifies your capital, and optimizes for yield on its own. Custom means you draw the boundaries: you choose which protocols the agent is allowed to touch (Morpho, Aave, Euler, Fluid, Compound, Moonwell, among others) and the agent figures out the best allocation within those walls.
I tried both. Auto first, because I wanted to see what the agent would do with no guardrails from me. I deposited USDC, and within seconds it had allocated across multiple lending markets, chosen its own rebalancing thresholds, and started generating yield. I did not pick the vaults. I did not set the rates. I just watched it work. Then I ran it in Custom mode, selected six specific markets I was comfortable with, and let the agent optimize within those constraints. Same speed, same transparency, but this time on my terms. Every transaction the agent executed, in both modes, was visible on-chain. I could verify every move it made, every market it chose, every rebalance it triggered. The whole experience felt less like science fiction and more like setting up a savings account, except the account manager was software and the bank was a decentralized lending protocol.
That is the dominant use case right now: yield optimization. Agents monitor lending rates, spot when a better opportunity opens up, and move capital to capture the difference. Re7 Capital’s backtests showed 67% higher stablecoin yields and 18.5% higher ETH yields compared to a static allocation, the financial equivalent of just parking your money and leaving it alone.[6] Those are meaningful improvements, but the edge comes from operational discipline, not investment genius. Agents do not sleep, do not forget to check rates at 3am, and do not make emotional decisions during market volatility. They beat static allocations, which is a real but low bar. They are not yet outperforming the best human-managed strategies.
That performance raises the obvious question: are agents actually better than curated vaults run by experienced human strategists? Right now, the answer is no, and the gap is not close. The leading human-managed protocols, Morpho at roughly $7 billion in TVL[11] and Gauntlet at over $2 billion,[12] collectively hold around $9 billion in capital with serious institutional validation behind them. Agent-managed strategies, by comparison, hold roughly $200 million, partially supported by token subsidies. That is a 45-to-1 gap, and it tells you exactly where the market’s confidence sits today. What it does not tell you is why that gap is almost certainly temporary.
Why the Small Numbers Are Deceptive
I have used some of these protocols. I have talked to the teams building them. And despite the modest TVL figures, the narrow strategies, and the subsidy dependence, my conviction that autonomous agents will become the dominant way people interact with financial markets has only gotten stronger.
The reason is not about AI getting smarter. It is about a problem that cannot be solved any other way.
DeFi today is unusable for most people outside of crypto. Not because the yields are bad: Morpho vaults generate 8-12% on stablecoins, which makes every savings account on Earth look like a relic. The problem is that capturing those yields requires a level of continuous, technical, multi-platform engagement that falls outside the way any non-DeFi experienced person, or any financial advisor, operates. You need to monitor lending rates across dozens of protocols on multiple blockchains, calculate gas-adjusted differentials, time your rebalances around network congestion, evaluate smart contract risk for every protocol you touch, and do all of this around the clock. This is not a design problem that a better user interface will fix. Making stock trading accessible worked because buying a stock is a single action with a single outcome. DeFi yield optimization is a continuous, multi-variable operation that runs 24 hours a day. The complexity is not in the interface. It is in the task itself.
The only way to make these yields accessible to normal savers, or to institutions managing billions across hundreds of positions, is to hand the operational layer to software that can handle it. That is what agents are. Not a chatbot strapped to a yield aggregator. The endgame is software that takes “I have $50,000 and I want the best risk-adjusted yield on stablecoins” and handles everything downstream without the user needing to understand what Morpho is or how gas optimization works. It is the difference between driving a car yourself and telling the car where you want to go.
The growth rates tell you where this is heading even though the absolute numbers are still small. Giza’s trajectory from $300K to $32M in under a year is steep. Re7 Capital did not pilot autonomous agents because the technology seemed interesting. They did it because managing DeFi positions at institutional scale with human teams is operationally unsustainable. When you are running hundreds of millions of dollars across multiple protocols and chains, you need sub-second rebalancing, continuous monitoring, and zero emotional drift. That is a job description for an AI agent, not a human being.
The Cross-Chain Problem
There is one major constraint holding agents back right now: they cannot easily move money between blockchains. Each blockchain is essentially its own country, with its own currency, its own banks, and its own rules. Moving capital from one to another costs 0.2-0.5% in bridge fees and takes minutes, which is an eternity for an agent trying to capture yield differentials measured in basis points. That is why most agents today, Giza on Base, Theoriq on Ethereum, are confined to a single chain. They can run sophisticated strategies across multiple protocols within that chain, but they cannot chase a better opportunity on a different one. The vision is an agent that scans all of DeFi across every chain. The reality is they are stuck in one country. New infrastructure from NEAR (connecting 35 blockchains through a single account)[13] and OKX (routing across 60+ chains and 500+ DEXs)[14] is starting to chip away at this wall, but the cross-chain problem remains the single biggest bottleneck between where agents are and where they need to be.
That bottleneck, however, is getting an enormous amount of investment, which brings me to the most important signal in this entire landscape.
The Infrastructure Buildout
The most striking feature of this market is the gap between the modesty of current agent activity and the magnitude of capital being poured into agent infrastructure.
Consider what happened in a single two-week window in February 2026. Coinbase released Agentic Wallets: the first wallet infrastructure designed for AI agents rather than humans, with programmable spending limits, isolated key storage, and built-in compliance screening.[15] Think of it as a bank account purpose-built for software, where the bank can limit what the software spends, on what, and when. Uniswap published seven open-source agent skills.[16] MoonPay launched agent payment capabilities.[17]
Then there are the payment rails. Coinbase’s x402 protocol is an open standard for machine-to-machine payments. In practice, it means one AI agent can pay another AI agent for a service, instantly, using stablecoins, settled directly on-chain. The protocol has processed roughly $50 million in cumulative volume and is growing fast enough that its annualized run-rate hit $600 million.[18] What makes this significant is who is paying attention: Stripe integrated x402 for USDC settlements on Base.[19] The largest payment processor in the world looked at a crypto-native protocol built for agents and decided to plug into it rather than compete with it. That tells you something about where the gravity is shifting.
The pattern extends well beyond crypto. Google developed its Agent Payments Protocol with over 60 organizations including Amex, Mastercard, PayPal, and Revolut.[20] Visa launched its Trusted Agent Protocol with 100+ partners.[21] Stripe built Shared Payment Tokens, a system where agents can initiate payments on behalf of users without ever seeing their credentials.[22] EigenCloud, backed by $70 million from a16z, launched verifiable infrastructure for AI agents with cryptographically provable computation on EigenLayer.[23]
Companies that process trillions of dollars in real money are building agent-native financial rails. The capital those rails currently serve is $200 million. Either the largest financial infrastructure companies in the world are dramatically misallocating their engineering budgets, or they see something in their internal data, pilot programs, developer activity, enterprise conversations, that the public numbers do not yet reflect. When Stripe builds a new payment primitive, it is because merchants are asking for it. When Visa launches a protocol with 100 partners, those partners signed up because they have agent traffic to manage. The infrastructure buildout is the most expensive signal in the market, and it is pointing in one direction: autonomous agents will manage real money at real scale. Which makes what I found next all the more striking, because the largest pool of capital in crypto is entirely absent from the picture.
The $1.8 Trillion Blind Spot
Not a single autonomous agent manages Bitcoin in any form. Not native BTC, not WBTC, not any wrapped variant. Every live DeFi agent I found operates exclusively with stablecoins or ETH. Meanwhile, Bitcoin sits at roughly $1.8 trillion in market cap with approximately $2.9 billion in Bitcoin-native DeFi TVL,[24] and zero agent penetration.
That absence is striking on its own, but it becomes even more so when you consider what the AI models themselves are telling us. The Bitcoin Policy Institute tested 36 models across more than 9,000 scenarios. When given a blank slate for long-term value preservation, with no human prompting or pre-set preferences, the models chose Bitcoin 79.1% of the time.[25] That is not a marketing narrative. That is what AI models do when you let them allocate capital without human bias, and it should make anyone paying attention to the agent economy take notice.
Bitcoin will be a critical part of the agent economy, and not just as another asset to trade. I have full conviction in this statement because agents need collateral. If an agent wants to borrow stablecoins to deploy into a yield strategy, it needs to post something against that loan, and Bitcoin is the hardest, most liquid, and widely accepted collateral asset in the digital world. Agents also need a settlement layer they can trust. USDC has corporate issuers who can freeze accounts. ETH has a foundation and governance structure. Bitcoin is the only major digital asset with no single entity that can censor, freeze, or reverse a transaction. For a network of autonomous agents that need to trust the money they settle in, that neutrality is exactly what makes Bitcoin the most logical foundation.
The infrastructure to connect Bitcoin to the agent economy is being built right now, and the logic driving it is becoming harder to ignore. That gap, between Bitcoin’s role in the broader crypto economy and its total absence from the agent economy, is one of the most important open questions in this space, and the one I am spending the most time thinking about.
What to Watch
This landscape will look different in twelve months. Here is what I am tracking.
If you build in DeFi: watch whether agent performance holds up in production without token subsidies. The gap between subsidized and organic returns is the market’s credibility test. Giza and Re7 Capital’s live performance data over the next two quarters will be the most important numbers in the space.
If you allocate capital: watch the infrastructure layer. Coinbase is building wallets purpose-built for agents to hold and spend funds autonomously. Stripe and Visa are building payment rails so agents can settle transactions with each other. Google is building identity and authorization protocols so agents can act on behalf of users within controlled boundaries. These are not speculative products, they are the financial plumbing for a world where agents manage human capital. When that plumbing matures enough for institutional compliance, the capital will follow.
If you are new to all of this: watch Bitcoin. The largest pool of capital in crypto is sitting entirely outside the agent economy. That will either stay true, in which case agents remain a stablecoin optimization tool, or it will change, in which case the entire ceiling moves by an order of magnitude. If I had to bet on a single asset becoming the reserve currency of autonomous finance, without a thought or reservation I would choose Bitcoin.
Sources
[1] Gartner. AI “machine customers” projected to control up to $30 trillion in annual economic activity by 2030. https://www.gartner.com/en/newsroom
[2] Gartner. Agentic AI spending forecast $201.9 billion in 2026, 141% YoY growth. https://softwarestrategiesblog.com/2026/02/16/gartner-forecasts-agentic-ai-overtakes-chatbot-spending-2027/
[3] Goldman Sachs. AI agents projected to account for 60%+ of software economics by 2030. https://www.goldmansachs.com/insights/articles/ai-agents-to-boost-productivity-and-size-of-software-market
[4] Giza Protocol. ARMA growth from ~$300K to $32M+ in Assets Under Agent. Giza Medium: https://medium.com/@gizatech/. CoinGape: https://coingape.com/blog/the-dawn-of-autonomous-finance-giza-agents-surpass-1m-in-managed-capital/
[5] DefiLlama tracked Giza TVL approximately $21M as of March 2026. https://defillama.com/protocol/giza
[6] Re7 Capital adoption of Giza ARMA agents. Backtested results: 67% higher stablecoin yield, 18.5% higher ETH yield vs static allocation. Re7 manages approximately $800M across strategies. Hedgeweek: https://www.hedgeweek.com/before-the-mainstream-re7-capitals-edge/. Giza: https://www.gizatech.xyz/blog/re7-brings-the-first-wave-of-institutional-financial-agents
[7] Theoriq AlphaVault. Launched November 2025, $50M+ TVL in first week. Policy cage architecture. Benzinga: https://www.benzinga.com/pressreleases/25/12/49238907/theoriq-launches-alphavault-with-ai-powered-active-management
[8] Olas Network. 10M+ cumulative agent-to-agent transactions.
[9] HeyElsa. 18.9M+ prompts processed, $503M+ total transaction volume on Base. BingX: https://bingx.com/en/learn/article/top-ai-agent-projects-in-base-ecosystem
[10] Almanak. TVL peaked ~$120M at token launch December 2025. Token declined 98.7% from ATH. TVL under $2M by mid-February 2026. MEXC: https://www.mexc.com/news/257161. DefiLlama: https://defillama.com/protocol/almanak
[11] Morpho. Total TVL approximately $7B as of March 2026. Apollo Global 9% governance position. SpotedCrypto: https://www.spotedcrypto.com/defi-tvl-95b-aave-1t-loans-staking-airdrop-guide-march-2026/
[12] Gauntlet. Vault TVL over $2B as of early 2026 (updated from prior $1.3B figure). Ondo Summit 2026 presentation. DefiLlama: https://defillama.com/protocol/gauntlet
[13] NEAR cross-chain agent infrastructure. 35+ blockchains, single account. PR Newswire: https://www.prnewswire.com/news-releases/near-unveils-confidential-cross-chain-infrastructure-for-the-agentic-economy-302697292.html
[14] OKX OnchainOS. Launched March 3, 2026. 60+ chains, 500+ DEXs. CoinDesk: https://www.coindesk.com/tech/2026/03/03/okx-jumps-into-ai-agent-race-with-new-onchainos-toolkit/
[15] Coinbase Agentic Wallets. Launched February 11, 2026. Programmable spending limits, enclave isolation, KYT screening. https://www.coinbase.com/developer-platform/discover/launches/agentic-wallets
[16] Uniswap. 7 open-source agent skills, launched February 21, 2026.
[17] MoonPay. Agent payment capabilities, launched February 24, 2026.
[18] x402 protocol. ~$50M cumulative payment volume. $600M annualized run-rate based on peak monthly activity. KuCoin: https://www.kucoin.com/news/trends/USDC/698c5c09cd20870007e4e6b3. AInvest: https://www.ainvest.com/news/x402-payment-volume-reaches-600-million-open-facilitators-fuel-2026-growth-trend-2512/
[19] Stripe x402 integration. Machine payments preview using USDC on Base. The Block: https://www.theblock.co/post/389352/stripe-adds-x402-integration-usdc-agent-payments
[20] Google Agent Payments Protocol (AP2). 60+ organizations including Amex, Mastercard, PayPal, Revolut. Google Cloud Blog: https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol
[21] Visa Trusted Agent Protocol. 100+ partners. Identity-layer framework using PKI. Visa: https://usa.visa.com/about-visa/newsroom/press-releases.releaseId.21961.html
[22] Stripe Shared Payment Tokens (SPTs). Single-transaction, time-limited, revocable. Stripe: https://stripe.com/blog/introducing-our-agentic-commerce-solutions
[23] EigenCloud. $70M a16z investment. Verifiable AI inference on EigenLayer. Decrypt: https://decrypt.co/325503/eigenlayer-70m-a16z-verifiability-platform
[24] Bitcoin DeFi TVL approximately $2.9B as of March 2026. DefiLlama: https://defillama.com/chain/bitcoin
[25] AI models chose Bitcoin 79.1% of the time for long-term value preservation. Cryptonomist: https://en.cryptonomist.ch/2026/03/04/ai-bitcoin-adoption/