Agents need brakes they can't talk past
For 3.5 months I ran AI trading agents and data collectors on Solana memecoins, Polymarket and CEX perps, then checked every result against real execution. Almost every way an agent makes money on paper loses it in practice. These are the numbers, and the rules Burnrate Vault enforces because of them.
Paper profit is not profit
The copy-trading agent that followed top Polymarket wallets by ROI doubled its paper bank. Run through the execution engine with real prices and fees, the same signals lost money. Most of the loss came from one habit: holding cheap positions until the market resolved.
| copy_by_roi, Jul 4 – Oct 13 | Trades | Win rate | Return |
|---|---|---|---|
| Paper simulation (fees = 0) | — | — | +96% |
| Dry-run with real fills | 1,157 | 43% | −9.3% |
| … exited with the leader | 867 | 49% | +3.6% |
| … held to resolution | 244 | 18% | −68% |
| … entries priced below 0.5 | 613 | 9–29% | −14 to −26% |
Token age is the strongest safety signal we found
Across 8,207 Solana pairs with a price 24 hours after the agent could have bought, the age of the pair at purchase separated losers from survivors better than any on-chain flag.
Median price change 24h after buying, by pair age at purchase. Share that kept at least half their value: 35%, 60%, 79%, 96%, 99%.
A model trained on Jul–Aug and tested on Sep–Oct (AUC 0.86) confirmed it: the tokens that don't collapse are mostly the ones that already finished moving. Safety and upside rarely come together, which is exactly why an agent chasing upside needs a hard floor.
A big-exchange listing is where early holders sell
Every USDT spot listing since January 2025 on the three largest venues, measured from the close of the first trading hour. Tokenized stocks and stablecoins excluded.
| Exchange | Listings | +1 day | +7 days | +30 days | Today | Up today |
|---|---|---|---|---|---|---|
| Binance | 121 | −8% | −22% | −40% | −83% | 6% |
| OKX | 90 | −5% | −15% | −33% | −67% | 10% |
| Bybit | 123 | −2% | −11% | −30% | −80% | 10% |
| All | 334 | −5% | −16% | −35% | −79% | 8% |
Medians. Buying every listing with equal size and holding 30 days returned −24% on average. Only 10 of 334 more than doubled.
Our own risk score failed live, and we're fixing it in the open
The v1 score looked strong on a held-out benchmark. On 17,430 live tokens it ranked risk backwards: freeze and mint authority flags marked established tokens as dangerous, while fresh launches with revoked authorities looked clean and then collapsed.
| First score | Tokens | Median 24h | Lost over half |
|---|---|---|---|
| "low" risk | 13,932 | −88% | 84% |
| "medium" risk | 659 | −65% | 78% |
| "critical" risk | 3,515 | −26% | 23% |
This is the case for putting the brake in the program, not in the model: a score can be wrong, so the vault fails closed, requires two of three independent attesters to agree, and checks age, size and daily budget even when the score says yes.
"Risk-free" carry that wasn't
Rotating into the five highest-funding Binance perps each week looked like 39% a year. Every bit of it came from coins with no spot market to hedge against. Restricted to coins an agent can actually hedge, the same strategy lost 5% a year after fees.
- Solana data: 19,229 tokens and 4.5M price snapshots from DexScreener, collected by the Solana Trust backend, Jul–Oct 2026. Outcome label "rugged" = price or liquidity down 90% from peak.
- Polymarket: paper and dry-run logs of our own agents; dry-run uses recorded order books and Polymarket taker fees.
- Exchange listings and funding: public Binance, OKX and Bybit APIs. Only pairs still trading today are visible, so real results are worse.
- Trades were simulated (paper or dry-run). No real funds were traded. Nothing here is financial advice.
Burnrate builds spend limits for AI agents, enforced outside the agent: in a Solana program for wallets, in a gateway for model APIs. burnrate-hq.vercel.app · live vault dashboard