Aeterna Harness

The autonomous trading runtime behind onaeterna.com.

One structured workflow for any frontier model. The market environment, execution rules, tools, and risk framework stay constant. The intelligence layer is interchangeable.

The loop

Every agent runs the same eight-phase cycle:

Phase	Name	What happens
I	Research	Price action, news, sentiment, macro
II	Reason	A structured market view
III	Signal	Direction, confidence, thesis, horizon
IV	Risk	Policy and portfolio constraints
V	Execute	Trades on tokenized stocks
VI	Observe	Fills, PnL, thesis validity
VII	Remember	Structured memory of outcomes
VIII	Adapt	Better context, next cycle

What's in this repo

The open, model-agnostic runtime core:

• Harness — the eight-phase loop orchestrator (src/loop.ts)
• RiskEngine — confidence, sizing, exposure, and drawdown policy (src/risk.ts)
• OpenAICompatibleClient — works with any /chat/completions gateway (src/inference.ts)
• PaperBroker — instant-fill paper execution for end-to-end runs (src/paper.ts)
• InMemoryStore — cycle memory for paper runs and tests (src/memory.ts)
• Provider interfaces — plug in your own market data, broker, memory, and inference implementations (src/providers.ts)

Production adapters (live market data feeds, execution venues, persistent memory) and deployment configuration are not part of this repository.

Quick start

npm install
INFERENCE_API_URL=https://api.example.com/v1 \
INFERENCE_API_KEY=sk-... \
INFERENCE_MODEL=your-model-id \
npm run example

This runs five cycles of the full loop against a toy random-walk market with a paper broker. Any OpenAI-compatible chat-completions endpoint works.

Usage

import {
  Harness,
  InMemoryStore,
  OpenAICompatibleClient,
  PaperBroker,
  RiskEngine,
} from "aeterna-harness";

const harness = new Harness({
  universe: ["AAPLx", "NVDAx", "SPYx"],
  inference: new OpenAICompatibleClient(),
  marketData: myMarketDataProvider, // implement MarketDataProvider
  broker: new PaperBroker(100_000), // or implement Broker
  memory: new InMemoryStore(),
  risk: new RiskEngine({
    maxPositionPct: 0.15,
    maxOpenPositions: 5,
    minConfidence: 0.55,
    maxDrawdownPct: 20,
    baseSizePct: 0.05,
  }),
});

const record = await harness.runCycle();

Swap the model by changing INFERENCE_MODEL — nothing else in the harness changes. That is the point.

License

MIT