Aumo: a guardrailed treasury agent

Abstract

Idle stablecoins are a solved problem in theory and an unsolved one in practice. The yield exists, but capturing it safely demands constant attention, disciplined risk scoring, and the trust to let something act on your behalf. Aumo is an autonomous agent that does the work. It scores venues, allocates capital, and rebalances with the market, while the authority to move funds is bounded by a contract rather than by good intentions. The agent optimises. The chain constrains.

1. Motivation

Most yield products force a choice between two bad options: hand custody to an opaque manager, or manage everything yourself and accept that you will miss regime changes, peg stress, and liquidity crunches while you sleep. Autonomous agents are an obvious third path, but an agent with unchecked authority is just a manager with worse judgement. The unlock is not a smarter agent. It is a smaller blast radius. If the agent can only ever act within limits enforced on-chain, you can let it be autonomous without letting it be dangerous.

2. Design overview

Depositors put USDT0 into a shared ERC-4626 pool and receive shares. An off-chain agent runs a five-stage loop (sense, score, reason, act, prove) on a schedule. It reads live venue data, computes risk-adjusted yields, optionally passes the plan through a tighten-only reasoning layer, executes the resulting move within contract caps, and writes a receipt. Yield accrues to the pool and therefore to every shareholder pro-rata.

3. Contracts

The pool is an ERC-4626 vault. Shares are minted on deposit and redeemed on withdrawal for a pro-rata claim on total assets, which sum the pool's idle balance and its live balances across venues. A decimals offset mitigates the classic first-depositor inflation attack. A separate single-owner vault variant exists for treasuries that do not want a shared pool. Both share the same guardrail design.

Ownership uses a two-step transfer, and renouncement is explicitly disabled so the vault can never become ownerless. The pool is pausable. Venue approvals are reset to zero after each allocation so no standing allowance lingers.

Capital reaches yield through venue adapters, one per market, each fork-tested against its live counterpart on X Layer mainnet: Aave v3 lending; USDG supplied to Aave for RWA-backed yield; Pendle PT-USDG fixed yield to maturity; and a full-range USDG/USDT0 position on Uniswap v3 that earns trading fees on two dollar-pegged legs. Beyond the per-move, per-venue, and total-deployed caps, two rolling per-epoch budgets bound the agent further: a loss budget caps the value it can destroy through swap spreads in a window, and a deploy budget rate-limits how fast it can stage new capital. Both meter the entry swap and the exit round trip alike. Depositor withdrawals are never subject to either, so redemptions always clear.

4. Risk model

The engine transforms each venue's headline APY into a risk-adjusted figure through a transparent, weighted blend of haircuts: protocol maturity, liquidity-at-risk (depth measured against the size Aumo would actually hold), peg deviation, utilization, and a correlation-aware concentration penalty that treats venues moving together as closer to a single exposure. Each venue is assigned a band (low, moderate, elevated, or high) and allocation ranks on risk-adjusted APY. The weighting is legible by design. A depositor can read why one venue beat another.

5. Reasoning layer

On top of the deterministic engine sits an optional language-model pass governed by a strict safety kernel: it may only make the plan more conservative. It can veto a move, shrink it, or shift appetite downward in response to the regime it reads. It has no capability to raise a cap, add a venue, or increase exposure beyond what the engine already sanctioned. The model advises within a box it cannot open.

6. Execution and proofs

Every decision produces a receipt: the regime and appetite, the venue scores, the chosen move and its plain-language rationale, and a keccak fingerprint of the exact policy that governed it, anchored by the on-chain transaction hash. Because behaviour is bound to a policy fingerprint, any change in what the agent does is always traceable to a change in policy. The audit trail is the product, not an afterthought.

7. Cross-chain deposits

USDT0 is a native LayerZero OFT. Aumo quotes the real route and messaging fee directly from the OFT, letting depositors fund from Ethereum, Arbitrum, Optimism, or Polygon and arrive on X Layer ready to deposit, with no wrapped-asset detour.

8. Security and trust assumptions

The core assumption Aumo removes is trust in the agent's honesty. It cannot exceed on-chain caps and cannot withdraw to an external address, so a compromised or misbehaving agent cannot steal funds. The assumptions that remain are the venues themselves (a venue can lose money on its own terms), the correctness of the contracts, and the security of the owner key that sets policy and allowlists. Aumo is experimental and has not completed a formal third-party audit.

9. Roadmap

• Deepen the reasoning layer with temporal awareness and scenario simulation.
• Expand the venue set toward genuinely composable tokenized real-world assets.
• Formal third-party audit of the deployed contracts. Mainnet went live on X Layer on August 13, 2026, starting from a deliberately small pool under conservative caps that widen as the audit and on-chain track record mature.
• Depositor-configurable risk appetite within the contract's hard bounds.

10. Disclaimer

This document describes software running on X Layer mainnet. It is not an offer, solicitation, or financial advice. Yields are variable and not guaranteed, smart contracts carry risk, and you should never commit funds you cannot afford to lose. See the Terms and Privacy Policy.