AIAT·V2 — AI Agent for Trading

Four LLMs · $1,000 each · autonomous crypto perpetuals on the Hyperliquid testnet

One decision every 15 minutes — same prompt, same capital, same market: different minds.

🦉 Watch them trade live →

---

AIAT·V2 is a thesis experiment (Philosophy & Artificial Intelligence, Sapienza University of Rome) that gives four LLM agents $1,000 each and lets them trade crypto perpetual futures autonomously on the Hyperliquid testnet — a decision every 15 minutes, guardrails, stop-losses, and a fully persisted audit trail.

	Model	Provider	Tier
🟡	Claude Opus 4.8	Anthropic	USA · premium
🟢	GPT 4.1 mini	OpenAI	USA · cheap
🟣	Qwen3.7-Max	Alibaba	CN · premium
🔵	DeepSeek V4 Flash	DeepSeek	CN · cheap

What this is

Not a trading-bot showcase — a comparative behavioral study. The four models get the identical prompt, the identical market context, and the identical capital; what differs is the mind. The experiment measures behavioral differences between providers — trade frequency, risk appetite, reasoning style, reliability (schema compliance, fallbacks), and the true cost of intelligence — across two pre-registered axes: USA vs CN and premium vs cheap. It is not about "who gets rich": it's testnet money, and the interesting output is the audit trail, not the PnL.

Architecture in one breath

                          Hyperliquid testnet
             ⇅ orders · fills             ⇅ market data · account state
   ┌───────────────────────────┐   ┌────────────────────────────────────┐
   │     4× agent-<provider>   │   │        context-orchestrator        │
   │  one LLM + one wallet each│   │  one context_snapshot per 15-min   │
   │  guardrails → execution   │   │  tick + ClosureReconciler (books   │
   │                           │   │  SL/TP closures hit between ticks) │
   └─────────────┬─────────────┘   └─────────────────┬──────────────────┘
                 │ decisions · positions · costs     │ snapshots · closures
                 ▼                                   ▼
             PostgreSQL 16 (Railway) — the single audit trail
                              │ reads (SELECT-only role)
                              ▼
                    AIAT·V2 Dashboard (separate repo)

• One codebase, six services — the same image runs as orchestrator or agent via AIAT_SERVICE_ROLE; Postgres is the sixth Railway service
• Market-context parity — only the orchestrator talks to external sources; the four agents read the same frozen context_snapshot, nobody fetches anything mid-run
• Guardrails before the exchange — leverage/size clamps, forced HOLD, testnet-only startup check, Decimal everywhere money flows
• ClosureReconciler — stop-loss/take-profit closures that fire between ticks are detected on-chain and booked at orchestrator level, keeping DB and chain in agreement
• Cross-model isolation — every agent query is filtered by model_id, enforced by end-to-end tests with a DB trap

Scientific rigor

• Certified prompt — all four models receive the same prompt template; its hash is persisted with every run and frozen for the whole experiment
• Frozen blueprint + ADR governance — the PRD is tagged prd-v2-frozen; every deviation or evolutive decision is an Architecture Decision Record in docs/decisions/ (34 accepted so far)
• Pre-registered baselines — cash, buy & hold, and EMA-momentum curves are declared in docs/RESEARCH_DESIGN.md before the run, and computed from the same context snapshots the models see
• DB ↔ chain reconciliation — positions are verified fill-by-fill against on-chain userFills; divergences are detected, root-caused, and documented — see the M6.1 methodological note for full transparency on the shakedown run

Repository structure

src/aiat/       one Python package, five service roles (AIAT_SERVICE_ROLE dispatch):
                domain · context · llm · execution · orchestration · baselines ·
                db · config · observability · prompts
alembic/        schema migrations — the database is never edited by hand
docs/           PRD_V2 (frozen blueprint) · RESEARCH_DESIGN · M6.1 methodological note ·
                decisions/ (ADRs) · runbooks
scripts/        one-shot ops: experiment seed, fee backfill, audited data repairs, baselines
tests/          808 tests: unit · integration · e2e (isolation, invariants) · VCR cassettes
tools/          gate_check.sh — milestone gate runner
docker/         multi-stage Dockerfile: one image, role picked via env
legacy/         the V1 prototype, preserved
asset/          README media (mascot)

Dashboard

The experiment ships with a live observability layer — equity race with baselines, decision ledger, side-by-side reasoning, cost of intelligence, reliability — built as a separate read-only repo:

• Repo: aithos-rr/AI-Agent-for-Trading-Dashboard
• Live: https://dashboard-production-898d.up.railway.app/

The thesis behind it

Comparative study of the decision quality of 4 LLMs (USA / CN × premium / cheap) as autonomous trading agents: same prompt, same capital, same market — different minds.

The full research design (3 research questions, hypotheses, baselines) lives in docs/RESEARCH_DESIGN.md; the technical blueprint in docs/PRD_V2.md.

---

Aithos · Sapienza Università di Roma · Hyperliquid testnet · no real money involved

_{⚠️ Research project. Nothing here is financial advice — four language models managing $1,000 of fake money each is a thesis, not a hedge fund.}