The three states of agent cash
Strip away the product names and an agent’s stablecoin balance is always in one of three places.
- Idle at 0%. Raw USDC in the agent’s wallet. Liquid, free of contract risk, unproductive. For most funded agents this is the default, because nobody wrote a sweep policy.
- Unregulated DeFi. Lending markets, LP positions and yield-agent products on Base, Arbitrum, Plasma and Solana. Real yield sometimes, incentive-driven yield often, and no license anywhere in the stack.
- x402 float. The working balance an agent keeps to pay for tools, inference and data over HTTP 402. Measured agent payments averaged ~$0.31 per transaction (Keyrock, May 2025–Apr 2026 window), so the float is small per agent and only matters in aggregate across a fleet.
Scale frames everything that follows. DefiLlama’s stablecoin float stood at ~$311.1B on 2026-09-04 (USDC $74.3B, USDT $184B), as read on our data panels. Cambrian’s Q2 2026 landscape counted labeled AgentFi TVL at roughly $12M on Base, $7M on Arbitrum and $417k on Ethereum — about $19M, or roughly 0.006% of the float. The two figures measure different objects (all pegged-dollar supply versus product TVL tagged as agent-managed), so the ratio is illustrative, not a market share. It still states the question this page answers: where does machine cash go once someone decides it should earn?
Agent Treasury Utilization: the metric
We define the flagship ratio as follows.
Agent Treasury Utilization = (vault + DeFi) / (cash + vault + DeFi)
cash = stablecoin balances in agent-labeled wallets, x402 float included
vault = regulated RWA vault shares, valued at convertToAssets()
DeFi = lending, LP and yield-agent positions, at redeemable value
Two sub-ratios travel with it: the regulated share, vault ÷ (vault + DeFi), and the float ratio, x402 float ÷ cash. A fleet holding 50,000 USDC in wallets, 30,000 in a yield agent and 20,000 in an RWA vault scores 50% utilization with a 40% regulated share. Utilization near zero is not automatically a failure. A solo x402 agent with $40 of float is rational to stay in cash, because one deposit transaction costs more than the yield it would earn. The number becomes meaningful at fleet and desk scale.
Honesty about the denominator: the numerator can be read on-chain today, sink by sink. The denominator needs an agent-wallet classifier with confidence scores, because no clean on-chain heuristic for “this wallet belongs to an agent” exists — candidates include ERC-8004-linked wallets, factory deployments, x402 payer addresses and EIP-7702 designators. So v0 publishes per-sink totals and a labeled-sample utilization, never a universe number. Every figure carries a window and a scrape timestamp, per our methodology.
DeFi yield agents: what exists and how big it is
The unregulated sink is a crowd of products that rotate stablecoins across lending venues and pools. The figures are Cambrian Q2 2026-era and will have moved; read them as order of magnitude.
| Product | Chains, as stated in our sources | AUM / TVL | What it does | Window · source |
|---|---|---|---|---|
| ZyFAI | Arbitrum · Base · Plasma | ~$3.9M | Stablecoin yield rotation | Q2 2026 · Cambrian |
| Yield Seeker | Base | ~$3.15M | Automated stablecoin yield | Q2 2026 · Cambrian |
| ARMA (Giza) · Almanak · Superform · Surf Liquid · AFI · Axal · Fraction AI Stable-Up | EVM | not broken out | Optimizers, vault routers, rebalancers; venues named across the category: Moonwell, Silo, Morpho, Yearn, Euler, Avantis | Q2 2026 · Cambrian list |
| 1tx-style mandate agents | EVM | n/a | Public policy diaries; reject unsustainable spot APYs | 2026 · public agent diaries |
| Kamino · Lulo | Solana | not broken out | Lending automation; yield aggregation | Q2 2026 · Cambrian list |
| Reflect | Base | not broken out | LP position management | Q2 2026 · Cambrian list |
| Labeled AgentFi TVL, total | Base · Arbitrum · Ethereum | ~$12M · ~$7M · ~$417k | Product TVL tagged as agent-managed | Q2 2026 · Cambrian |
Three observations. First, the largest products hold three to four million dollars each — smaller than one mid-sized DeFi whale. These are pilots, not treasuries. Second, every hop a yield agent takes adds a smart contract, and much of the APY it chases is reflexive: paid in token incentives or by borrowers levering the same asset, and gone when the price turns. The 1tx pattern — a public mandate that refuses spot APYs it cannot underwrite — is the first sign of discipline in the category. Third, Cambrian’s observation holds in production: rule-based agents still dominate when capital is at risk, and the working pattern is “LLM proposes, policy engine disposes”. Which wallet enforces that policy is the subject of our provider comparison.
The regulated path: IXS Agentic Vault, read on-chain
IXS sells the alternative an institution can approve: humans and autonomous agents use the same vaults, and the vaults wrap regulated real-world assets rather than crypto-reflexive loops. The platform describes itself as licensed under the Bahamas DARE Act, cites Coinbase Ventures, UOB Venture Management and Spartan Group as backers, and shows “$88M+ raised on platform”, “60+ institutional deals” and seven years of operating history (ixs.finance, fetched 2026-09-04). Our research notes record the ~$88M figure as TVL independently cited via RWA.io; we have not re-verified that listing, so we quote both wordings rather than pick one.
The IXS Agentic Vault was announced live on 2026-07-09 with Finance District. The mechanics: an agent deposits USDC into on-chain yield sourced from institutional RWAs; no human signs each deposit inside policy; BNB Chain first, with the H2 2026 roadmap (published 2026-07-16) naming Base as “coming next”; a published, verifiable contract at 0xc975a3EeF2e49F8eDdEf585340C43f15300fCB82; ERC-4626 shares; REST, SDK and MCP access with x402-compatible APIs; endpoints listed in x402, Circle’s agent directory, agentic.market and the CryptoSkill MCP directory. Access runs through the Finance District Agent Wallet — EVM, Solana and Bitcoin wallets provisioned at signup, keys in a TEE, non-custodial — and Compass Labs turned the sentence “Deposit 500 USDC into the IXS vault on BNB Chain” into an MCP payload. IXS cites capital deployment in under 60 seconds. The roadmap describes the live BNB Chain vault as wrapping BlackRock high-yield corporate bond exposure.
- ~4% — Fidelity USD money market fund exposure. Permissionless: any EVM wallet, no protocol-layer KYC, wallet-level AML at the contract.
- ~6% — BlackRock corporate bond exposure. Permissionless; the profile the roadmap attaches to the live BNB Chain vault.
- ~9% — private credit. Permissioned: KYC of about three minutes, offered “across Asia and select jurisdictions”.
- 4–12% — BTC real yield, paid in stablecoins without selling BTC. Permissioned.
All four are indicative APYs, exactly as IXS displayed them on 2026-09-04. They move. Core vaults carry no lock-up; IXS describes redemption as “exit on demand”. FDUSD and further stablecoins are announced to follow.
We do not take scale on faith. On 2026-09-04 we called totalAssets() and totalSupply() on the published contract over a public BNB Chain RPC: 4,347.09 USDC of total assets (18-decimal BNB Chain representation; shown as $4,347 on our data page), against 3,993.66 vault shares (symbol ixv1), or 1.0885 assets per share. Read plainly: two months after launch, the flagship regulated sink for agent treasuries held about four thousand dollars. That is a pilot-sized deposit and the honest baseline for the chart we will keep drawing. A share price above 1.0 is consistent with accrual since launch, but we have not verified the opening exchange rate, so we do not publish an implied yield from it.
How to evaluate a sink
Licensed vault or yield agent, the same nine questions apply. The graph stores the answers as vault attributes — license, kyc, chain, rail — so an allocator can separate regulated sinks from farms in one query.
| Check | The question | IXS Agentic Vault, as published | Typical DeFi yield agent |
|---|---|---|---|
| License flag | Who is licensed, where, for what? | Bahamas DARE (IXS-cited) | None in the stack |
| KYC flag | Required at which layer? | Permissionless for treasury and bond products; KYC for private credit and BTC yield | None; the wallet is the identity |
| Lock-up | Can the agent exit today? | None on core vaults; “exit on demand” | Varies by venue; LP exits depend on pool depth |
| Daily liquidity | Redemption size versus vault size? | Untested at scale: $4,347 total assets on 2026-09-04 | Pool-dependent; large exits move the rate |
| Underlying | What is actually earning? | MMF, corporate bond, private credit, BTC structures | Lending spreads, LP fees, token incentives |
| Chain | Where does the position live? | BNB Chain; Base next per roadmap | Base, Arbitrum, Plasma, Solana |
| x402 / MCP access | Can an agent discover and deposit without a human? | MCP, x402-compatible APIs, directory listings | Mixed; many products are app-first |
| Smart-contract risk | Audits, upgradeability, admin keys? | Contract published; not reviewed by us | Stacked: every routed venue adds a contract |
| Reflexive yield | Does the APY survive a price drop? | RWA-sourced, lower ceiling | Often incentive- or leverage-driven |
ERC-4626 makes the first pass cheap: one interface — deposit, redeem, totalAssets, convertToAssets — prices shares across every compliant vault, so an agent can compare sinks without bespoke code. It defines the interface only. The standard says nothing about what stands behind the shares, which is the point of the other eight rows.
A decision table by operator type
This is research framing, not investment advice. It describes where cash tends to sit given each operator’s binding constraint, not what any operator should buy.
| Operator | Typical idle balance | Binding constraint | Where the cash usually belongs |
|---|---|---|---|
| Solo builder, one x402 agent | $10–$500 | Fees exceed yield | Cash. Float sized to a few days of API spend; utilization near 0 is rational |
| Fleet operator, 10–1,000 agents on one master wallet | $5k–$500k | Sweep policy and daily liquidity | Allowance per agent in cash; excess swept to a permissionless ERC-4626 vault or a mandate-style yield agent with no lock-up |
| Fund or treasury desk with a written mandate | $500k+ | License, underlying disclosure, audit trail | Regulated RWA vaults, permissioned tiers where KYC is available; DeFi only inside explicit policy limits |
| Regulated institution piloting agents | Pilot-sized | Licensed venues only; human approval above a threshold | Licensed vault plus a policy engine that escalates on size; the menu is thin today, and that is the finding |
| Solana-native fleet | Any | No regulated ERC-4626 path in our sources | Kamino or Lulo automation; the regulated path is EVM-only for now |
Whatever the row, the sweep is a treasury action and belongs inside the wallet’s policy engine — a spend limit, an allowlisted vault address, a size threshold that escalates to a human. Delegation through EIP-7702 session keys is what makes “no human signature per deposit” safe rather than reckless; our wallet-connection guide walks through the setup.
Risks and the regulatory perimeter
The perimeter does not close. MiCA regulates issuers and service providers in the EU, the GENIUS Act regulates dollar-stablecoin issuers in the US, and the EU AI Act regulates AI systems by risk class. Each touches a piece. None cleanly covers a machine paying a machine, or a machine allocating a treasury with no human signature per transaction. “Licensed” therefore describes the issuer and the underlying, never the agent’s decision — which is why the graph flags licensed vaults and farms separately instead of letting one word cover both.
The liability vacuum is real. Ledger documented a 2026 case in which an attacker manipulated an agent into transferring ~$175k through a hidden, Morse-code-style instruction — not a stolen key, not a broken AMM. A yield mandate widens that surface: “move idle cash to the best available APY” is precisely the sentence an injected page will try to redefine. Allowlisted sink addresses and size thresholds are the countermeasure; our security checklist lists the rest.
Contract and liquidity risk scale differently. Q1 2026 DeFi hacks exceeded $1B, and a routed yield agent inherits the risk of every venue it touches. A regulated vault concentrates risk in one contract and one issuer instead — cleaner, but a $4,347 vault has never processed a $1M redemption, so “daily liquidity” is a claim until tested. Concentration compounds it: USDC settles ~99% of measured agent payments, and a sink that only takes USDC inherits that monoculture. Indicative APYs move, and nothing on this page is investment advice. What we can promise is the accounting: the vault’s total assets, deposit and redeem cadence, and the discovery path each depositing agent took, on the graph and the data panels. The wider context sits in what agentic finance is and the research hub.
Sources
- IXS permissionless vaults page (indicative APYs, no lock-up, ERC-4626, $88M+ raised, 60+ deals) — fetched 2026-09-04 — link
- IXS H2 2026 roadmap (BNB Chain live with Finance District, Base next, directory listings, ERC-4626 native) — published 2026-07-16, fetched 2026-09-04 — link
- IXS vault interface (client-rendered; heading only without JavaScript) — fetched 2026-09-04 — link
- Agentic Finance Graph on-chain read:
totalAssets()andtotalSupply()on 0xc975…CB82, BNB Chain, public RPC — 2026-09-04 — link - DefiLlama stablecoin float, USDC and USDT circulating — snapshot 2026-09-04 via our panels — link
- Cambrian Agentic Finance Landscape, Q2 2026 (AgentFi TVL by chain; ZyFAI and Yield Seeker AUM) — mid-2026, as tabulated on our data page — link
- Keyrock, “Who Pays the Agent” (average ticket ~$0.31; ~99% USDC) — May 2025–Apr 2026 window — link
- RWA.xyz — IXS TVL independently cited per our research notes, which also reference RWA.io; listing not re-verified by us — 2026 — link
FAQ
What is Agent Treasury Utilization?
The share of agent-held stablecoins that is working: (vault + DeFi) divided by (cash + vault + DeFi). Cash includes x402 float. Vault means regulated RWA vault shares valued at convertToAssets; DeFi means lending, LP and yield-agent positions at redeemable value. We publish it per labeled sample with a window and a scrape timestamp, not as a universe number.
Why does ERC-4626 matter for AI agents?
ERC-4626 standardizes deposit, redeem, totalAssets and convertToAssets, so one integration prices shares across every compliant vault and an agent can compare sinks without bespoke code. It defines the interface only. It says nothing about the underlying, the license, the admin keys or the audit, which is why the evaluation checklist exists.
How much is in the IXS Agentic Vault right now?
Our own eth_call to totalAssets() on the published BNB Chain contract returned about 4,347 USDC on 2026-09-04, against 3,993.66 vault shares. The live value is on our data page. The indicative APYs IXS publishes, about 4% to 12% depending on product, are marketing figures that move. This is research, not investment advice.
Can an AI agent deposit into a yield vault without a human signing each transaction?
Yes, inside a policy. The Finance District Agent Wallet keeps keys in a trusted execution environment, any MCP-speaking agent can talk to it, and the operator sets the mandate once. The vault does not need a per-deposit signature; the policy engine decides whether the deposit is allowed. The human is replaced by the policy, not removed from the system.
Should an agent keep any USDC idle?
Usually yes. x402 tickets averaged about 0.31 dollars in the Keyrock window, so a few days of API spend is a small float, and moving it into a sink costs more in fees and risk than it earns. Utilization near zero is rational for a single small agent; it becomes a problem at fleet and treasury scale.
Do MiCA, the GENIUS Act or the EU AI Act cover machine-to-machine payments?
Not cleanly. Each regulates a piece: stablecoin issuers, reserves, AI systems. None addresses a machine allocating a treasury without a human signature per transaction. Licensed, in this context, describes the issuer and the underlying asset, not the agent decision, which is why the graph flags licensed vaults and farms separately.