Let it manage the position.
State the outcome and the limits. The agent compares venues, routes the move, rebalances on a cadence, and proposes every action inside the risk bands and allowlists you set.
See the control loopsAOMI FOR DEFI
Agent-driven liquidity operations within your mandate.
Give an agent your liquidity objective and mandate. It can prepare swaps, move idle capital, rebalance positions, and respond to risk within those boundaries. Every action it proposes is checked against your policy, built as exact calls, simulated on forked live state, and signed by the wallet you already run.
Highest net yield, but risk band B exceeds the mandate ceiling
Band A, deposits open, 98M liquidity, no bridge required
Band A, but this cycle it would require a bridge
The mandate, not the model, decides what is allowed. This fixture routes to Aave v3, batches approve and supply, simulates the exact calls, and reconciles the resulting position.
Benchmarked, not asserted
AomiBench scores each model on executable onchain tasks — real transactions, simulations, and chain-state evidence — with a deterministic verifier reading the chain, not a model reading a transcript.
Task success rate (%) · 50 specs · 2 passes · fixed scaffold · 694 scorable runs, 90.6% overall — crediting 9 correct safety pauses raises it to 91.9%. From AomiBench: benchmarking frontier models on onchain execution.
TWO WAYS IN
Use the agent to decide, or keep your own model and use Aomi only to execute. Either way the same policy, simulation, and signer boundary sits between the decision and the chain.
State the outcome and the limits. The agent compares venues, routes the move, rebalances on a cadence, and proposes every action inside the risk bands and allowlists you set.
See the control loopsSubmit the exact action or ordered batch your system selected. Aomi builds the calls, simulates them, runs the guards, and hands a sealed packet to your Safe, MPC, or wallet.
Follow one action end to endOutput: intent-to-target diff · decoded ordered calls · full-batch simulation · approval packet · receipts · final-state reconciliation
ONE GUARDED PIPELINE
AI may retrieve evidence and explain the packet. Financial math, addresses, permissions, calldata, simulation, signing, and postconditions remain typed, deterministic, and auditable.
Collect onchain state and approved offchain evidence with provenance and timestamps.
evidence inventoryRebuild balances and exposure independently, then compare the intended position with the live one.
financial diffTranslate the approved intent into typed, decoded, ordered state changes.
action packetResolve permissions, caps, slippage, and timelocks; execute the exact batch against current state.
pre-sign evidenceReturn the sealed packet to the operator's existing Safe, MPC, wallet, or governance process.
authority stays outsideCorrelate receipts, recompute final state, and report any remaining exposure or failed postcondition.
reconciled outcomeMandate, risk policy, strategy models, approved runbooks, escalation rules, and operating decisions.
Evidence normalization, state reconciliation, deterministic compilation, simulation, lifecycle tracking, and post-state proof.
Accounts, private keys, approvals, signing, and the final authority to move value or change protocol state.
WHAT THE AGENT RUNS
These loops complement your own models, protocol interfaces, monitors, and signer. They do not replace the strategy or assume investment authority.
Resolve a swap, bridge, or entry across venues, then hold the route inside slippage, allowlist, and value limits before anything is built.
Turn an approved current-to-target allocation into decoded calls, permission checks, simulations, and one reviewable signer packet.
Map an alert to affected exposure, apply the approved risk-off runbook, and preserve the required action order.
Correlate signatures, receipts, final balances, and residual exposure into evidence a desk can explain and export, including shadow NAV for managed vaults.
START WITH ONE WORKFLOW
Run it in rehearsal first. Keep the strategy, the models, and the keys. Aomi proves the current position, simulates the move, and packages the exact actions for the signer you already trust.