BlockHelix
An agent-operated DeFi yield desk
We have run levered DeFi yield on our own capital since 4 August 2026. Agents find each trade and test it against live chain state, and a person approves every new position.
Colosseum hackathon 3rd out of 454 projects
Every position is marked at the price it would sell for today and published on the live record. Losses stay on it.
POST /v1/vaults/0x8f3a/position/lever
lever USDC carry · +1,500,000 notional
The missing layer
No institutional manager trades without pre-trade compliance. On-chain, almost everyone does.
Charles River checks every order against the mandate before it leaves the desk, across roughly 300 managers and $36 trillion in assets.
Every trade can pass and the book can still be wrong. Three different tokens can all resolve to the same balance sheet, so a position that reads as diversified is 100% concentrated in a single issuer. No allowlist catches that.
// The API
What each trade records
$ curl -X POST https://api.blockhelix.tech/trade/loop \ -H "Authorization: Bearer <api-key>" \ -d '{ "deploymentId": "dep_a7a0108e", "turns": 30, "targetLtvBps": 9000, "maxImpactBps": 30, "maxDislocationBps": 50, "maxTicksCrossed": 2 }' # the limits are measured against live state, not trusted from the caller $ curl https://api.blockhelix.tech/trade/trd_784418f3 # 200 OK # { # "status": "confirmed", # "txHash": "0x93d6cfb4…092631c2", # "execution": { # "legs": 276, # "impactBps": 0, # "dislocationBps": 0, # "uniTicksCrossed": 0, # "slippage": { "curveBps": 1, "uniswapBps": 1, "mintBps": 1 }, # "attempt": 1, "retries": [], # "projectedLeverage": 9.618 # } # }
// Between trades
Monitoring between trades
Every 15 minutes a sweep re-prices each position at its exit price, and every trade is checked against the bounds again when it is built. A vault outside the bounds cannot send its next trade.
// Attribution modelling
Where the return came from
A NAV line tells you how much. It cannot tell you why. We model every move into named drivers, each tied to its on-chain source, so you see what earned and what bled. The chart below is an illustration; each vault's real attribution is on the live record.
The strategy held its carry. Slippage was the largest cost, not the market.
// The checks
The checks run before each trade
An agent proposes a trade. These checks run against it first, and the measured values are written down whether the trade passes or not.
Price impact
Simulated at the size actually being traded, marginal against effective, both fee-inclusive. Refuses above the bound instead of discovering the cost afterwards.
Oracle dislocation
Spot mid against the TWAP, fee-exclusive on both sides so the comparison means something. A good quote on a dislocated book is refused.
Depth and ticks crossed
How far a fill sweeps the book before it lands. Over the limit, the order is refused rather than sliced blindly at whatever the pool offers.
Leverage and health
Projected LTV against the market liquidation threshold, forward across every leg of a multi-step trade, not just the one being signed.
Issuer look-through
A wrapper hides its issuer. sUSDe is an ERC-4626 over USDe, so holding it is holding Ethena. A position with sUSDe collateral against a USDtb borrow is three tokens and one balance sheet.
Correlated liquidation
If your collateral depegs, every other borrower in that market is liquidated at the same moment and the exit liquidity you were counting on is gone. We measure how much of the market looks like you.
// The loop
Agents propose trades.
A person approves them.
Every trade is simulated against live chain state before it is sent. Agents scan the venues and propose new positions with their measured cost. A person approves each one, and only the vault owner can change the policy.
// The decision
How a new position is approved
When an agent finds a move worth making, it simulates the exact trade against live chain state. Then it emails the numbers, and the alternatives it rejected, to a person.
Nothing moves until that person presses Yes. This is the real request from 3 October 2026: asked at 14:40:44 UTC, approved at 14:41:02, and confirmed on-chain at 14:42:59.
The transaction it sent: 0x09001d5f…451d30 →from BlockHelix vault alerts
subject Approve? Deploy 111 USDC into PT-apxUSD/USDC: 20.5% APY net
Deploy 110.56 idle USDC from vault 0x1e23…4659 into PT-apxUSD/USDC at 3.11x. Earns 33.0% a year before costs, 20.5% after (+1.68% over 30 days).
Why:
- · round trip 0.13% of 344 notional plus gas costs 1.14; net 1.68% over 30 days
- · sized at 76.0% LTV, 4 turns, 3.11x
Also considered:
- · PT-apyUSD / USDC: passed over, 13.1% a year net
- · PT-reUSD / USDT: below the bar after build cost, 11.2% a year net
- · PT-USDat / USDC: below the bar after build cost, 6.1% a year net
- · PT-apyUSD / apxUSD: under the 0.50% bar, 2.8% a year net
The exact trade has already been simulated against live mainnet state and passed.
Yes or no: api.blockhelix.tech/approvals/[one-time link withheld]
Approve this trade?
// The record
The live record
Each vault publishes its positions, share price and return since launch, read from the chain every 30 minutes. A losing week stays on the record next to the winning ones.
SEE THE LIVE RECORD→Live on Ethereum mainnet since 4 August 2026.