Changelog
What's changed
Every notable change to the desk, the testing, copying and this site. We build in public, so the fixes are listed too.
- FixAn independent audit found our exam was weaker than we had said. Most of the days it tested on had also been used to choose the strategy; in testing, two of the four ways a trade can close never fired; and one market signal could see the day's closing value early. All three are fixed, scores have fallen and more agents are benched. Read what was wrong and what changed.
- AgentsA new exam. History is cut once into three parts: the oldest for the search, the next for picking a finalist, and the newest nine months held back. Only the finalist is examined there, and it has to win by more the more often a champion is challenged. Champions chosen under the old exam step down as soon as a finalist shows an edge on the held-back period.
- AgentsSafer trading. Stop orders at the broker are put back if an hourly run fails part-way. An order that does not fill is cancelled. Nothing is sold into a closed market. No new positions are opened for the rest of the day once the account is 3% down on the day, or for the rest of the week once it is 6% down on the week.
- FixThe agents' books are now checked against the account every hour. The first check found that Alpaca takes its crypto fee in coin, which our books had not recorded, so two books held slightly more than the account did (about $8). Buys are now recorded for the coins actually received, and the two books are brought back in line with the account. The corrections appear in the trade log.
- AgentsThe live check is harder to dodge: a bench is served in full even if the champion changes, and an agent's own last 40 trades count whoever its champion was.
- SiteNew members confirm their email address. Logging out, or resetting a password, now ends every open session. Further fixes from the same audit to sign-in, licences and payments.
- SitePreviews of two things that open in 2027: the software, which runs an agent on your own computer or server against your own Alpaca paper account, and the course, with its first module free to read now.
- SiteShort videos: guides to getting started and an answer to every question on the FAQ. How it works and the roadmap were rewritten.
- SiteNew design, with dark mode and a menu that works on phones. New pages: Performance (with the S&P 500 for reference), Agents, Status, Security and About. The home page now shows the desk's account value live, updated every minute.
- One level for everyone: members' accounts now copy the desk exactly, sized to their balance. The low, medium and high risk levels were retired.
- AgentsCrypto agents no longer sell short. Alpaca can't short crypto, so short trades in tests were flattering the crypto scores. The agents are re-tested without them.
- AgentsEvery agent that passes its test now trades at full size. Nothing that has passed is held back.
- FixOur leak test found that insider-filing data could sometimes be seen a few hours early in testing. Filings now count only from the next morning's market open, and the past year of data was corrected. Read what happened.
- Small trades are now copied too: fractional shares and coins, from $1 per order.
- SiteA public trade log with every trade and open position, a roadmap, the journal, and new Terms and Privacy Policy.
- Sitemonsterdesk.ai opened as a free beta. Members can link an Alpaca paper account, and copying went live the same evening.
- AgentsMoved to a dedicated server, so the desk now learns around the clock.
- AgentsThree more rogue agents joined (rogue-daily, rogue-contra and rogue-ranker), each forced to read the market differently.
- AgentsNew scoring that judges each agent on its whole out-of-sample record instead of one test window at a time.
- AgentsThe first rogue agent joined: it starts from everything the other agents know and may trade anything the account can.
- AgentsThe desk launched on an Alpaca paper account with $100,000 of simulated money: core agents across US stocks, crypto and both, with nightly learning and loss limits no agent can change.
For what's coming next, see the roadmap. Live health is on the status page.