Part of Nous Ergon — Autonomous Multi-Agent Trading System. Repo and S3 names use the underlying project name
alpha-engine.
Multi-agent investment-research pipeline. Six sector teams, a CIO, and a macro economist scan the S&P 500+400 weekly, maintain rolling investment theses, and emit signals.json for the rest of the system. Built on LangGraph with Anthropic Claude (Haiku per-team, Sonnet for synthesis).
System overview, Step Function orchestration, and module relationships live in
alpha-engine-docs. Code tour and key files live inOVERVIEW.md.
- Six sector teams (Tech, Healthcare, Financials, Industrials, Consumer, Defensives) run in parallel via LangGraph
Send()fan-out. Each team: quant ReAct → qual ReAct (with RAG retrieval over SEC filings + earnings + theses) → peer review → 2–3 ranked recommendations + thesis update. - Macro economist runs in parallel with a reflection loop, producing market regime + per-sector ratings that scale recommendations downstream.
- CIO evaluates every recommendation in a single Sonnet batch call against a 4-dimension rubric and gates new entrants per a configurable cap.
- LLM-as-judge layer scores agent outputs at key stages against rubric prompts. Every decision is captured to S3 with prompt metadata + cost telemetry for replay and audit.
This is where every agent decision in the system happens. Each one is captured as a structured artifact — prompt id + version + hash, full prompt context, input snapshot, agent output, and cost — replayable, auditable, and attributable to a specific prompt revision. The LLM-as-judge layer scores agent quality at key stages against rubric prompts. Together these are the substrate that lets Phase 3 measure whether prompt or model changes actually improve agent quality.
flowchart LR
Data[Data + RAG<br/>universe · macro · filings] --> Teams[6 sector teams<br/>quant → qual → peer review]
Data --> Macro[Macro economist<br/>regime + sector ratings]
Teams --> CIO[CIO synthesis<br/>Sonnet · 4-dim rubric<br/>entrant gate]
Macro --> CIO
CIO --> Out[signals.json<br/>thesis history<br/>decision artifacts]
Decision-artifact capture wraps every LLM call site via LoadedPrompt (frontmatter-versioned prompts, sha256 body hash) and a track_llm_cost ContextVar accumulator that stamps token counts + cost on each invocation.
This repo is public. Agent prompts, scoring weights, universe configuration, and proprietary scoring formulas are gitignored locally and stored in the private alpha-engine-config repo. Architecture and approach are public; specific values are private.
| Module | Repo |
|---|---|
| Executor | alpha-engine |
| Data | alpha-engine-data |
| Predictor | alpha-engine-predictor |
| Backtester | alpha-engine-backtester |
| Dashboard | alpha-engine-dashboard |
| Library | alpha-engine-lib |
| Docs | alpha-engine-docs |
MIT — see LICENSE.