Robinhood Agentic

Ask a language model about a stock and it answers. It answers with identical confidence whether or not it has read a filing, a price, or a headline. That failure mode is what this project exists to make structurally impossible.

Grounding

The research loop reads real holdings over a read-only Robinhood connection, prices from yfinance, news from RSS, filings from SEC EDGAR, and a persistent memory of what it concluded before. Sixteen typed tools sit behind a hand-written Claude tool-use loop with a twelve-step budget. Enough steps to screen, inspect, retrieve, and log. Few enough that a wandering run terminates instead of drifting.

It never trades

That is a decision, not a missing feature. An agent that can both form a thesis and act on it deletes the one checkpoint that catches a bad thesis. Automating the click was never the valuable part.

Knowing whether it works

Researcher, analyst, monitor, and judge workflows run on FastAPI and asyncio with cooldowns, failure isolation, persistent state, and streaming audit traces.

On top sits an evaluation harness: LLM-as-judge scoring, human failure labels, one-pass repair gates, and a shadow portfolio benchmarked against SPY and sector indexes. A research agent that cannot beat the index it implicitly claims to beat is not a research agent.