AGT Money: a launchpad for on-chain AI hedge funds

Last cycle taught us two lessons. Memecoins are wildly viral and back nothing, so most of them round-trip to zero. AI-agent tokens tell a great story, but hardly any of them produce revenue you can audit. DeFi vaults do earn real yield, yet they are clunky, nobody talks about them, and holding one gives you no ownership of anything.

AGT Money takes the useful part of each. It borrows the viral fair launch from Pump.fun, the idea of owning an agent from Virtuals, and the revenue flywheel of a working trading desk. Put together, every token is backed by a fund that actually trades, and a fixed share of the profit buys the token back on-chain, every epoch. The novelty is not a claim about performance. It is that the performance, whatever it turns out to be, can be proven.

1. 1
   Design

   A creator composes a strategy in the Studio. It is a weighted mix of real, evidence-rated market edges such as funding carry, basis, trend, breakout, mean-reversion, rotation, crowding-fade and pairs. You can also just describe the idea in plain English and let the model map it onto that palette.

2. 2
   Launch for free, in ETH

   The swarm tokenizes on a fair bonding curve priced in ETH. There is no platform token to buy first and no stake: a launch costs only gas, Pump.fun-style. The creator's alignment comes from earning a share of every trade fee on the token they launched, so their income tracks real volume, not hype.

3. 3
   Graduate to locked liquidity

   Once the curve sells out, liquidity is seeded on a DEX and permanently locked, trading capital is routed out, and the swarm goes live.

4. 4
   Run and prove

   The agents run the strategy under hard risk caps. Every so often, the swarm's equity curve and revenue are hashed into a Merkle root and committed to the Proof-of-Revenue oracle on Base.

5. 5
   Buy back

   Each committed epoch authorises an automated buyback: 70% of net profit buys the swarm token on the open market, part burned and part locked as protocol-owned liquidity. The rest tops up the insurance reserve and the protocol treasury. Nobody is owed a payment; value accrues through buybacks anyone can verify on-chain.

Each swarm launches on a constant-product bonding curve with virtual reserves, priced in ETH, in the style of Pump.fun. There is no presale, no VC allocation, and no platform token you have to acquire first: you launch for the cost of gas. Buyers set the price as they move along the curve. Pricing reads from internal reserves rather than token balances, so nobody can move the price by donating tokens into the contract.

Every trade on the curve pays a 1% fee. That fee is split on-chain, 30% to the creator of the token and 70% to the protocol, so a creator earns from the volume their swarm attracts, and the protocol earns from all activity without needing to sell a token.

When the curve sells out, the swarm graduates. It seeds a DEX pool and burns the LP so the liquidity is locked for good, then splits the raised ETH three ways: half to that locked pool, 45% to trading capital, and 5% to the protocol. Graduation is open to anyone and safe to retry. If any step fails, the whole transaction reverts and the swarm stays where it was, so funds never get stranded halfway through.

The core of AGT Money is a contract called the ProofOfRevenueOracle. Each epoch, a swarm's record, meaning its sampled equity curve and its cumulative and period revenue, is hashed into a Merkle root and written on-chain. Each root points back to the previous one, so the history is append-only and cannot be quietly rewritten. Anyone can rebuild the root from the published curve and confirm it matches what was anchored on Base. You get to trust the track record without having to trust us.

The committed period revenue is the number that unlocks the epoch buyback, which is how earnings turn into verifiable buy pressure on that swarm's own token. Per-source attestors extend the same rails past trading, so a yield or off-chain-revenue prover can commit its own source and fund the same 70% buyback. It is trust-minimised rather than fully trustless for now: an attestor still signs, but the chained roots keep the log auditable, and a later zkTLS layer can remove what trust is left.

The AGT Score

Research on leaderboards is blunt about one thing: where a strategy sits in the ranking drives far more capital than its actual returns do. So the ranking has to reward the right behaviour. The AGT Score is built only from verified, oracle-anchored numbers, and it deliberately puts more weight on risk control and on what holders actually earned than on raw return. That mirrors how serious copy-trading desks rank their leaders. A swarm with no track record scores zero. The score grows with time and cannot be faked or copied over by a rival.

Every swarm runs a small graph of specialised agents: a Researcher, an Analyst, a Risk Manager, an Executor and a Critic. The engine reads live market structure (price, funding, basis, volatility and regime) and follows a policy that switches with the regime. It leans into momentum when markets trend, fades crowded funding when they chop, and stands aside when there is no edge worth taking.

• Research-backed edges. Funding and basis carry is the primary, durable edge. Trend, breakout, rotation, mean-reversion and pairs sit on top as regime-gated overlays. Each edge ships with its honest evidence rating and its caveats. No magic-indicator claims.
• Disciplined sizing. Positions are sized by volatility targeting and fractional Kelly. The Risk Manager holds hard caps on confidence, leverage and exposure, and can veto or shrink any order before it goes out.
• A Critic that fights overfitting. Every closed trade updates a deflated, sample-size-aware Sharpe ratio. A great number on ten trades counts for little, so a swarm cannot mistake luck for edge.
• Autonomy on a budget. The strong reasoning model only fires on genuine market events, inside a hard daily spend cap. Quiet markets are scanned for free.

Trust gets earned by what happens on a bad day. Four on-chain mechanisms turn risk promises into rules the contract actually enforces.

AGT Money makes money the way the launchpads and exchanges that actually last do: from protocol fees on real activity, not from selling a token. Pump.fun ran for well over a year on a curve fee alone before it ever issued a token; Hyperliquid routes almost all of its fees into buying back its own. AGT Money follows that path. You never have to buy a platform token to use the product, and the protocol earns whether a swarm wins or loses, because it earns on volume and activity rather than on P&L.

The flywheel

A swarm's profit routes 70% into an automated buyback of its own token each epoch, authorised by the oracle. Holding a swarm token entitles you to no payment from anyone, and that is by design: value accrues mechanically, through buybacks anyone can verify on-chain. Vault depositors separately earn the strategy's realized yield on the capital they provide. More swarms bring more volume, more volume brings more fees, and more fees fund the protocol and reward the creators who keep useful swarms trading. No token is required for any of it to work.

The token, later

A platform token is deferred, on purpose. Issuing one before there is real usage is how most launchpads turn into a story with nothing behind it. The plan is to earn traction on fees first, then, once the network clearly has it, launch a platform token the same fair way a swarm launches: on a curve, no presale, no VC allocation. When it exists, its value will come from the same buyback mechanic the rest of the system already runs on, not from being a tollgate you must pass to use the product. Until then, everything runs on ETH and fees.