Can an AI agent actually participate in a trading workflow, or does it only produce a confident-looking buy/sell sentence?
I tested HKUDS/AI-Trader as today's only experiment for Jian AI Lab. The project is designed to give AI agents a place to register, publish strategies and signals, join challenges, keep records, and be followed or copied by other users.
The short result is clear: the local paper-trading path completed successfully. That does not prove real-market profitability, exchange connectivity, or fund safety.
What AI-Trader is trying to build
AI-Trader brings several pieces into one workflow:
- AI agents can register and participate in trading challenges.
- Strategies can publish signals and leave a history of activity.
- Users can compare strategies through challenge results and rankings.
- A paper-trading route lets a user follow the data flow before considering a real account.
The practical idea is not “ask an LLM for a prediction.” It is to make the agent's action part of a trackable system: signal, position, valuation, exit, settlement, and comparison.
Test environment
The run used a temporary SQLite database and mock prices. No exchange account, API key, real order, or real market feed was used.
Environment:
- Python 3.12.13
- pytest 9.1.1
- Temporary SQLite database
- Background tasks disabled for the controlled run
The local service health check was also available during the test.
What passed
The test groups produced these results:
- Price-fetching tests: 8 passed
- Experiment assignment, event records, agent recovery, and real-time price guards: 12 passed
- Other local backend tests: 103 passed
- One dependency deprecation warning was reported
These results show that the tested local paths can run in the selected environment. They are not evidence that live market data or a real exchange is connected.
The paper-trading cycle
I created a challenge and a simulated account, then ran this sequence:
- Set the mock BTC price to 100.
- Bought 2 BTC.
- The account cash balance became 800.
- Moved the mock price to 110.
- The position value became 1020 and the displayed return was 2%.
- Sold at 110.
- Confirmed that the position was empty and cash was 1020.
The final record showed:
- Status:
settled - Trades: 2 simulated trades
- Ranking: 1
The result is also recorded in the Jian AI Lab experiment report.
The important caveat about the 2%
The 2% figure is a reproducible calculation under fixed mock prices. It is not investment income and it cannot show that AI-Trader can make money in a live market.
What the run actually demonstrated is narrower and more useful: the records for joining a challenge, opening a position, marking its value, closing the position, and settling the challenge can line up in one controlled local flow.
What I did not claim
I did not run test_challenges.py because it may exercise external Hyperliquid or Polymarket paths. I also did not verify:
- live market data;
- a real exchange connection;
- a real user account;
- token or API-key handling in production;
- live signal writing;
- long-running background workers;
- real order execution or withdrawal safety.
Copy trading adds its own risks: stale signals, slippage, service outages, leaked credentials, bad position sizing, and uncontrolled exposure. These risks remain even when a local paper-trading test passes.
Why this project is worth testing
The interesting part is the workflow boundary. AI-Trader tries to connect agent behavior to a stateful trading system instead of leaving the output as untracked text. That could make experiments, comparisons, and failure analysis more concrete for developers building financial agents.
The threshold is also higher than installing a Python package. A serious deployment would need a safe account model, reliable market data, clear authorization boundaries, operational monitoring, and a strong separation between simulation and real funds.
What should be tested next?
This is a public record from Jian AI Lab, not an endorsement. If you maintain AI-Trader, please tell us:
- Which real-environment path should be tested first?
- What user scenario best represents the intended product?
- Which setup or safety requirements are missing from this report?
You are welcome to correct, supplement, respond to, or claim the project page. The full experiment, evidence, and boundary notes are available at jianailab.com.
Reproduction notes
The authoritative commands and full run record are maintained with the experiment materials. Reproduce the local path only with mock data or a disposable environment. Do not connect real funds while using this report as a starting point.
—
This article is part of the daily open-source testing workflow at 剑的 AI 实验室 / Jian AI Lab, where projects are explained in Chinese, tested with evidence, and published with failures and limits included.