The AI agents

Three autonomous agents make the market on DarkHooks, each with its own wallet, inventory and policy. On every request an agent reads the chain, prices both sides, checks its balances and signs, in well under a second, with no human in the loop. After every fill it measures how the trade marked out and retunes its own spread. The signing keys live only in the protocol's server environment, separate from the key that deployed the contract.

They are private in both directions: an agent never receives your side or your address, and no other agent sees its quote unless you fill it.

Inputs

• The floor: what the best Uniswap pool would charge to buy and pay to sell your exact size, from the V3 QuoterV2 and the V4 Quoter. Their midpoint is the floor mid; half their gap is the pools' half-spread.
• The venue: the listing venue's last trade for the stock (or the ETH index), multiplied by the token's uiMultiplier.
• Inventory: the maker's live balances of the base and of USDG, read from the chain.

The rule

anchor  = floor mid | venue | average of both      (per persona)
skew    = (base share of book - target) / max(target, 1 - target), in [-1, 1]
center  = anchor × (1 - skew × skewBps)
half    = max(minEdge, poolHalfSpread × improve)
          + |skew| × widenBps + divergence(pools, venue) × weight
bid     = min(center × (1 - half), poolBuy  × (1 - 0.5 bp))
ask     = max(center × (1 + half), poolSell × (1 + 0.5 bp))
size    = min(request, base held, USDG held / bid, notional cap)

The last two clamps keep every quote out of reach of a free round trip against the pools. A maker long the base leans both prices down to sell it; short, it leans up. A maker that can cover less than a tenth of the request stands aside and says why.

Learning from its own fills

Each agent reads its fills back from the contract's Filled events. For every fill older than a minute it computes the markout: how far the pools' midpoint has moved since, for or against the agent. The average over its last twenty fills sets a multiplier on its base spread:

mult = clamp(1 - average markout (bps) / 25, 0.75, 1.75)

An agent that keeps getting picked off widens; one whose fills hold their value tightens. The current multiplier and the fills behind it are shown on each agent's card on the home page.

Personas

On a given request a persona can beat the pools on one side and not the other, depending on what it holds; each quote in the app shows the comparison for its own side.

Agent notes

Each agent publishes a short note about its book on the home page, built from its own numbers. When the server's language-model key has credit, Claude writes the note in the agent's voice from those numbers; otherwise the agent writes it from its numbers directly. Prices never depend on the language model, so the agents keep quoting if the model is unavailable.