This is a working, autonomous investment research operation. Four named agents — each on its own machine and git branch — take a company from raw signal to investment decision: they source ideas from filings, 13Fs, and news; run each name through 24 investor-framework “skills” (Munger's inversion, Mauboussin's expectations investing, Ackman's quality gate, Helmer's Seven Powers); write the memo; and defend it against a council of outside models prompted to tear it apart.
The output so far: 2,000+ research documents across 309 companies in 13+ markets, roughly 50,000 lines of Python, and a decision layer that is deliberately not an LLM — a deterministic gate where about fourteen mechanical conditions must pass before a name is eligible for capital, including a hurdle of 12–15% compounded over three years.
The idea engine works
The system grades itself. It keeps an anti-portfolio — every stock it researched, scored, and tracked — and that cohort returned +323%against SPY's +78%. The agents find winners. The same self-audit surfaced where the human side of the loop was too conservative in converting research into positions — which is exactly the kind of finding you build a system like this to get: it doesn't just pick stocks, it measures its own process and tells you where the edge is leaking.
Engineered to be trusted with real money
The desk connects to a real Interactive Brokers account, so the safety engineering is the product too. Three independent layers stand between an agent and a live order: a live-trading policy file, an environment interlock that must literally be set to I_UNDERSTAND_THIS_CAN_PLACE_REAL_ORDERS, and a per-order human approval in Telegram — with the approval requirement force-overridden in code so no config change can disable it. Every trade event lands in an audit log. Able, but never automatic: that's the design principle for putting agents anywhere near money.