Agentic Trader

The short version

A sentiment-driven multi-agent trading system: bulk news ingestion — RSS, Reddit, SEC EDGAR — feeds agents that maintain a running narrative state per symbol and produce portfolio-aware trade proposals against a market simulator. Everything is verified end-to-end: a replay simulator, a live-mirror simulator, and 890+ tests pinning the behavior, capped with a full 30-day simulation replay.

What I built

• The ingestion layer. Bulk news and filing feeds (RSS, Reddit, SEC EDGAR) normalized into per-symbol narrative state that agents can reason over.
• The proposal loop. Sentiment agents propose trades that are portfolio-aware — sized and filtered against what the book already holds, not in isolation.
• The simulation stack. A market simulator, a replay simulator for historical runs, and a live-mirror mode; the whole pipeline was verified against Alpaca paper trading.
• The discipline. ADR-driven development, 890+ passing tests, and a full 30-day simulation replay as the acceptance run.

What was hard

Not the strategies — the plumbing. Idempotency on every side effect, clock discipline so replay and live modes behave identically, and keeping the simulation honest instead of accidentally profitable. Most of the real work was making lies hard to tell myself.

What happened next

The 30-day replay ran clean end-to-end. It stays in paper mode until the approach survives substantially more history; the system is built so that turning up the replay window is a configuration change, not a rewrite.

What I took from it

Verification-first beats backtesting vibes. A result you can replay, re-run, and pin with tests is worth more than a good month of paper P&L — and replaying history honestly is the hardest, most valuable part of the stack.