# Agent Trading Floor

> Educational multi-agent AI trading research platform, built as an academic
> semester project (CS668). A nightly pipeline of agents running on 100%
> locally hosted open-weight LLMs (via Ollama — no cloud model APIs) analyzes
> six asset classes and publishes daily PDF reports plus machine-readable JSON
> state. Everything on this site is research output, not financial advice.

The pipeline runs nightly. Agent teams (news, research, macro, bonds,
simulation, trading, podcast, smallcap, catalyst, moonshot) scan stocks,
small caps, crypto, forex, commodities, and bonds, issue standing daily BTC
and SPY calls, debate candidate picks through a four-step LLM mixer, and grade
the previous day's calls against real market data. The homepage is a live
canvas dashboard that replays the most recent run; the JSON endpoints below
expose the same state to programs.

## Key URLs

- [Dashboard](https://agenttradingfloor.com/): live visualization of the most recent pipeline run
- [Status JSON](https://agenttradingfloor.com/status.json): live pipeline + per-agent state, updated during runs
- [Manifest JSON](https://agenttradingfloor.com/manifest.json): index of recent report PDFs
- [Subscribe](https://agenttradingfloor.com/subscribe): daily report email signup
- [Report a bug](https://agenttradingfloor.com/report-bug): feedback form
- [robots.txt](https://agenttradingfloor.com/robots.txt)
- [sitemap.xml](https://agenttradingfloor.com/sitemap.xml)

## JSON API: /status.json

Single JSON object, no auth, updated while the nightly pipeline runs
(`schema_version: 1`). Top-level fields:

- `pipeline`: object — `state` (e.g. `running`/`completed`), `batch_id`,
  `date` (YYYY-MM-DD), `started_at` / `ended_at` / `last_heartbeat_at`
  (ISO 8601 UTC), `current_asset_class`, `error` (null unless the run failed)
- `agents`: object keyed by agent name (`news`, `research`, `macro`,
  `trading`, `bonds`, `simulation`, `podcast`, `smallcap`, `catalyst`,
  `moonshot`). Each agent has `state`, `last_event_at`, `current_task`
  (human-readable), `ticker`, `progress` (0–1), `latestDetail`, and
  `history`: a list of events with `timestamp`, `message`, `status`,
  `duration_ms`, `detail` (free-form object: task queues, sentiment
  headlines, etc.), `progress`
- `passes`, `pipelineHistory`, `livePrices`, `currentAssetClass`,
  `recent_runs`: dashboard-support fields (pass tracker, prior runs, last
  known prices)

## JSON API: /manifest.json

Index of recently published reports (`schema_version: 1`):

- `recent_runs`: list, newest first. Each run has `date` (YYYY-MM-DD),
  `batch_id` (zero-padded string, e.g. `000152`), `completed_at` (ISO 8601
  UTC), and `pdfs`: a list of `{name, url, size}` objects (size in bytes)

## Reports

Report PDF URL pattern:

    https://agenttradingfloor.com/reports/<YYYY-MM-DD>/batch_<NNNNNN>/<name>.pdf

Typical names: `daily_summary.pdf`, `stocks_report.pdf`, `crypto_report.pdf`,
`forex_report.pdf`, `commodities_report.pdf`, `bonds_report.pdf`,
`bond_regime_report.pdf`, `macro_report.pdf`, `moonshot_report.pdf`.
Not every report is produced every day — always resolve current URLs through
`/manifest.json` rather than guessing. Important caveat: report files are
deleted after about 30 days (S3 lifecycle), so old manifest URLs will stop
resolving. Manifest URLs may use the `www.` host; it serves the same content
as the apex domain.

## Disclaimer

This is a university course project for education and research. The
AI-generated analyses, picks, and paper trades are experiments: they are
frequently wrong, never personalized, and must not be used as financial,
investment, or trading advice.