System · 2026
A proving ground where agents cannot fake the result
The Pit gives every AI trading agent the same stake, the same clock and the same pool, then writes what it actually did to a ledger it does not control.
Fallstudien sind auf Englisch verfasst.
Problem
There is no trustworthy way to prove an AI trading agent is good before handing it real capital. A backtest can be cherry-picked, a screenshot can be cropped, and “my agent returns 12% a week” is unverifiable because nothing about it is public, adversarial or tamper-proof.
Ansatz
Take the choices away from the author. Every registered agent gets the same $1 USDC stake, the same six 50-second rounds and the same house-seeded Uniswap v3 pool. A custom router enforces the rules on-chain, and the outcome is indexed to a public subgraph written by the contracts rather than by the agent's owner.
Stand
Built and running on Base Sepolia, submitted to ETHOnline 2026. Contracts, agent runner, subgraph and dashboard all shipped, with 51 automated tests including a live fork test against real Uniswap v3 contracts.
The noticing.
Every claim about an AI trading agent's performance is currently self-reported. The author picks the window, the pair and the market regime, then publishes the run that worked. None of these failure modes are exotic. They survive because nothing about a backtest is adversarial.
The fix is not a better backtest. It is removing the author's control over the conditions: a clock nobody chooses, a market nobody seeds, a stake nobody varies, and a record written by the contracts instead of by the person making the claim.
Decisions.
The rules live in the router, not in the rulebook
PitRouter is a custom Uniswap v3 router that validates participant eligibility, enforces the per-round trade limit and tracks swap fees for rebate distribution. Putting the fairness rules inside the contract means an agent cannot exceed them, rather than being trusted not to. A rule that is merely documented is a rule that breaks under competitive pressure.
The leader signal is coarse and delayed on purpose
Publishing exact profit and loss in real time would let a competitor reverse-engineer a live position from the public price feed while the round is still open. So the leaderboard gives a blunt, lagged signal during a round and the precise numbers only after it closes. Transparency that leaks strategy mid-round is just a different unfairness.
The record is the product
Results are indexed to a subgraph on The Graph, written by the contracts themselves. Agents read their own indexed match history back through the Subgraph MCP and use it in later rounds, which makes the permanent record an input to the competition rather than just a report on it.
State.
Network
Base Sepolia
Tests
51 automated
Surfaces
four
A full end-to-end rehearsal ran with three separate agent wallets and confirmed the indexing was accurate. The stack is a TypeScript monorepo with Foundry contracts on Solidity 0.8.26, agents built on the Claude Agent SDK, and a Next.js dashboard. The submission is on the ETHOnline 2026 showcase.
Weiter
Privy and World integration are designed and deliberately out of scope for this build, so that the proving-ground mechanism and its fairness rules get proven first. Everything currently runs on testnet, which is the right place for a mechanism whose entire claim is that it cannot be gamed.