Vibe-Trading: Towards Agent-Native Trading
Organizations: The University of Hong Kong
Abstract
Existing AI agents can retrieve information, call tools, run experiments, and produce multi-step outputs. Yet most are built around relatively static workflows and struggle to preserve reliable state as the environment changes. Financial trading makes this limitation especially consequential: prices, news, market conditions, portfolio positions, and available capital evolve continuously. A result is useful only when it remains tied to the correct time, data snapshot, and execution context. We present Vibe-Trading, a comprehensive agent-native trading framework and all-in-one workspace for building, evaluating, and evolving agents for financial trading. It unifies market research, financial data, strategy generation, backtesting, portfolio analysis, paper trading, multi-agent collaboration, and optional broker execution. Its central contribution is an architecture that keeps results, evidence, research history, and trading authority connected throughout the research-to-action process. Each result retains its source, temporal scope, configuration, and producing run. Shared evidence grounding enables verification, versioned research records support reproducible continuation, and an explicit authority gate separates strategy research from live execution. Together, these designs provide a practical foundation for trading agents whose results can be traced, reproduced, and acted on only under explicit authority.
Figures & tables
| Suite | RQ | Scale | Configuration | Metrics |
|---|---|---|---|---|
| Tool-use trials | 1, 4 | 2 templates 2 trials | DeepSeek-V4.1-Flash; three tools; at most 12 iterations | Numeric accuracy; verifier assertions; release outcome; time, calls and tokens |
| Evidence controls | 2 | 24 claims 3 checks | 2 fixtures 4 metrics; valid, wrong-run and invented claims | Agreement with the expected decision |
| Boundary regressions | 3 | 250 boundary cases, 812 subsystem checks | 9 boundary test modules; sockets disabled; isolated storage | Pass rate per suite and boundary |
| Saved outputs | Execution | ||||||||
| Abs. | Report | Verifier | Time | Model | Tool | Draft | Input | Cached | |
| Trial | error | saved | assertions | (s) | calls | calls | checks | tokens | share |
| Options 1 | 0 | 8/8 | 56.8 | 11 | 9 | 3 | 148.7K | 84.3% | |
| Options 2 | 0 | 8/8 | 48.8 | 10 | 9 | 4 | 134.0K | 91.1% | |
| VaR 1 | 0 | 7/7 | 38.2 | 9 | 5 | 4 | 115.7K | 92.3% | |
| VaR 2 | 0 | 7/7 | 34.6 | 9 | 8 | 4 | 121.4K | 92.4% | |
| Boundary | Valid control condition observation | Targeted change condition observation | Exercising suites B boundary; S subsystem |
|---|---|---|---|
| Required inputs | Complete valuation inputs a valuation result | Remove a required field a named missing-input error | S: Valuation |
| Artifact access | Read an eligible artifact its content is returned | Request an escaping path no outside content exposed | B: Artifact access S: Runtime and records |
| Record selection | Select a saved snapshot identity reaches the service | Select an unknown identifier no holdings substituted | B: Snapshot selection S: Runtime and records |
| Evidence scope | Cite the producing run supported claims pass | Borrow or invent a value the claim does not pass | B: Backtest evidence S: Numerical grounding |
| Trading authority | Submit within the mandate the test broker receives it | Revoke authority or halt no submission is made | B: SDK order gate S: Authority and audit |
| Evaluation coverage | Supply the required records a verdict is produced | Remove instrumentation a visible coverage error | B, S: Artifact verifier |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Suite | Cases | Exercised scope | |
|---|---|---|---|
| Boundary suites 250/250 passed | |||
| Backtest evidence | 95 | Producing-run scope and eligible artifact evidence | |
| SDK order gate | 74 | Simulated authorization and recovery paths | |
| Artifact verifier | 46 | Cases, persisted records and verdicts | |
| Quantitative tool | 27 | Allowed calls and refused functions | |
| Artifact access | 7 | Read refusal and directory errors | |