Catalyst's AI Trading Agents Can Trade Without Reviewing Each Order
Sequoia led a $30 million seed for agents that turn plain-language instructions into leveraged crypto trades. Catalyst's terms allow trading "without your contemporaneous review", making the scope of each user's permission central to the risk.
Catalyst says its trading-agent pilot generated hundreds of millions of dollars in volume over a couple of weeks. The users were power traders. The figure comes from the company and was generated by agents that sign trades for users based on plain-English instructions.
Sequoia led Catalyst's $30 million seed round. Jump Trading, PeakXV, Lux Capital, AntiFund, Coinbase and Premji Invest also participated, Fortune reported. Fortune says "a human in the loop has to accept". Co-founder Dylan Iskandar says the goal is that "people will simply state an intent, and we have our agents to figure out everything end to end."
The risk depends on what the user accepts. Catalyst's homepage shows a demo prompt: if the Associated Press tweets positively about Anthropic, capitalise on it; if not, short it; if CNN tweets something opposite, sell the position; and take profit and set a stop loss along the way.
A user who accepts that instruction once leaves it active until two news accounts satisfy it, possibly while the user is away. The page does not indicate that the trading venue sees the trigger. The venue most likely receives only the signed order.
Catalyst's terms of service explain how it works. Each transaction goes through a non-custodial wallet built on Privy infrastructure. Catalyst says it does not take custody of the assets.
Users "may delegate signing permissions to an AI Agent, allowing the AI Agent to sign and submit Transactions on your behalf without requiring your approval at the time of each Transaction". Limits can apply to the transaction type, amount, recipient or smart contract. Transactions outside those limits will not execute.
The terms also say agents "may execute Transactions rapidly, repeatedly, and without your contemporaneous review". Users may incur losses "before you are able to detect or stop them".
An instruction such as "Capitalise on it" gives the signer little detail. It sets no position size, leverage or time limit, leaving the agent to choose within the delegated permission. That makes the model's interpretation a secondary concern: a misreading can spend only what the delegation allows, but a correct reading of a bad instruction can spend it just as quickly.
The user grants signing rights with a ceiling. Catalyst's FAQ says users set the budget, strategy and execution conditions, and the system stays within those boundaries. The terms list possible limits but leave their values to the user.
The FAQ says users can pause or modify strategies at any time, and the terms allow them to revoke permissions. Those controls help only if someone is watching when a trigger fires. A model's interpretation can be checked by asking again; a ceiling set too high becomes visible only when a trade reaches it.
The FAQ lists spot tokens, perpetuals, tokenized stocks and prediction markets, with more onchain and offchain assets planned. Users can deposit with Apple Pay, a credit card or USDC, SOL, ETH and BNB. For now, withdrawals are in stablecoins.
Perpetual contracts use leverage. A stop-loss instruction is another standing order executed by the same agent under the same permission. In a fast market, prices may move past the stop before it executes. It limits losses only if the trigger works as the user expects. When a position closes at a loss, the stablecoin balance left is whatever the permission allowed the agent to retain.
The terms call outputs informational and make the user solely responsible for reviewing instructions before execution. That places most of the review burden on the moment permission is granted, even though users are also told to check account activity.
The clearest consequence may fall on trading venues. Many users can create strategies that watch the same tweet. If it appears, each strategy could fire within moments from a different delegated signing address. Each order may stay within its own limit and look ordinary on its own.
A venue's risk engine would see one correlated burst spread across many small accounts. Margin rules, open-interest limits and per-address position caps assume traders act independently. A shared public trigger breaks that assumption.
Venues could classify agent-signed orders separately and ask wallet providers to expose permission limits before an order arrives. Wallet providers could make amount, contract and expiry limits the default, with users choosing to widen them. That would slow onboarding for the speed-focused customers in Catalyst's pilot, but establish a limit before a loss occurs.
Fortune says Catalyst is now admitting people from its waitlist and expects wider availability later. The FAQ says the product is free during the research preview.
The first published prices will show which customers Catalyst is targeting. A per-trade fee rewards higher trading volume and may encourage broader permissions and faster onboarding. A flat monthly fee rewards user growth without tying revenue to how much each agent trades.
The pilot attracted power traders, who may pay for volume. Fortune says the broader goal is retail trading for people who pay for convenience and may not watch each trade. When Catalyst publishes its execution fees, a per-trade schedule would point to volume as a priority; a flat fee would leave more of the risk control with the wallet user.