Research notes on portfolio intelligence, risk models, and the future of AI-assisted investing
An investing agent can follow every formal rule and still misunderstand the user. We built a second agent to review its plan and completed work.
Natural language is how you explore an investment idea. A formal, machine-checked contract is how that idea becomes something the agent must prove.
Two unedited, public sessions — a stress-tested tactical strategy and a fossil-free index tracker — with the reasoning, checks, and repairs left visible.
We asked it to find a tradable signal. After 29 rigorous steps it said there wasn't one — and that refusal is the point.
Fundamental analysis and quant optimization each answer half the question. Here's how the agent runs both at once, worked through GOOGL vs AAPL.
The process and math behind Itoflow's in-house risk model for global stocks, ETFs, style tilts, mixed-asset covariance, and single-stock shocks.
Hooks can't scale. Sub-agents block. We built asynchronous critique agents that use spaced repetition to course-correct long-running AI trading strategies without ever stalling the main execution.
When AI agents decide to rebalance at 10 PM, naive systems execute at stale prices. Here's how we built async execution that queues trades and waits for each exchange to open.
The patterns that made Claude Code successful don't transfer to investing. Here's why the challenges are fundamentally different, and harder in ways that aren't immediately obvious.