Vibe-Trading: Towards Agent-Native Trading

Haozhe Wu, Zhonghang Li, Chao HuangarXiv 2026 · arXiv

TL;DR

Vibe-Trading is an "agent-native" framework designed to solve the problem of state drift and lack of provenance in AI-driven financial trading. By strictly linking every result to its specific data snapshot, temporal scope, and producing run, it ensures that trading decisions are grounded in verifiable evidence and executed only under explicit, audited authority.

What Problem Does This Paper Address?

Most current AI agent frameworks are built for static workflows. They struggle in dynamic environments like financial markets, where prices, news, and portfolio positions change by the second. In this context, a numerical result (like a "Buy" signal) is only useful if it remains tied to the exact moment and data snapshot that produced it.

Without rigorous state tracking, agents suffer from "state drift"—using outdated observations for new actions—or "hallucinated evidence," where they report numbers that don't actually exist in their research history. Furthermore, the risk of an autonomous agent executing unauthorized or duplicate trades poses a significant financial and operational threat. Vibe-Trading addresses these gaps by creating a unified environment where research, backtesting, and execution are architecturally connected but logically separated by an "authority gate."

Methods

The authors introduce an agent-native architecture, meaning that financial results, supporting evidence, and action permissions are explicit objects that agents can inspect and use. The framework consists of two cooperating layers:

  1. The Financial Research Workspace: This is the "engine room." It provides data access, backtesting, and analytics. Crucially, it enforces Financial Computation Contracts, ensuring that every result carries its interpretation basis (currency, units, timing) and that backtests are reproducible by freezing all conditions except the strategy logic.
  2. The Financial Agent Harness: This layer acts as the "supervisor." It manages the agent's context (memory, skills, and token budget) and controls how the agent interacts with the workspace.

A key innovation is the Evidence-Grounded Execution cycle. Before an agent can release a report or propose an action, the harness performs Scoped Validation. It checks the agent's claims against an "admitted evidence set"—a ledger of results actually produced by the workspace tools. This prevents the agent from "borrowing" values from the wrong research run or inventing them entirely.

Figure 1: Overview of Vibe-Trading. The workspace provides financial computation, research records and authorized execution; the harness coordinates policy decisions, tool execution and report validation.

Datasets

The paper evaluates the framework using three distinct suites:

  • Tool-use trials: Using synthetic inputs for Option Pricing (Black-Scholes) and Historical Value-at-Risk (VaR) calculations.
  • Evidence controls: A set of 24 claims (valid, wrong-run, and invented) used to test the system's ability to detect fraudulent or erroneous reporting.
  • Boundary regressions: 250 test cases designed to ensure the system's security boundaries (like the authority gate and artifact access) cannot be bypassed.

The evaluation utilized the DeepSeek-V4.1-Flash model as the agent policy.

Key Insights and Ideas

  • Identity-Preserving Evidence: The "aha moment" here is that a number is not just a value; it's a triplet of (Value, Source, Context). By treating research artifacts as immutable, versioned records, Vibe-Trading allows agents to compare strategies across different market regimes without losing track of which data led to which conclusion.
  • Separation of Research and Authority: The framework treats "knowing what to do" and "having permission to do it" as entirely separate. Even a perfectly backtested strategy cannot execute a trade unless it passes through an explicit Authority Gate that checks against user-defined mandates (e.g., leverage limits, instrument restrictions).
  • Layered Diagnosis: The authors found that agents often succeed at the "artifact level" (calculating the right number) but fail at the "delivery level" (reporting it to the user with the correct units or semantics). This highlights the need for automated verifiers that sit between the agent's internal reasoning and its final output.
Figure 2: One tool-call cycle in the harness: each policy request is executed under the relevant controls, and its result or error informs the next decision.

Key Results

  • Numeric Precision: In end-to-end trials, the agents achieved zero absolute error compared to independent mathematical oracles for both option pricing and VaR tasks.
  • Robust Validation: The "Full Ledger" validation mechanism successfully accepted all 8 valid claims and rejected all 16 "wrong-run" or "invented" claims. In contrast, simpler "value-only" checks failed to catch agents borrowing numbers from the wrong research sessions.
  • Operational Efficiency: Trials typically completed within 60 seconds, with the harness successfully using prompt caching for 84% to 92% of input tokens, making the research process both fast and cost-effective.
  • Boundary Integrity: The system passed 100% of the 250 boundary regression cases, confirming that the controls effectively block prohibited transitions (like unauthorized trading or out-of-scope data access).
Figure 3: Execution timelines of the four tool-use trials.

Limitations and Future Work

While Vibe-Trading provides a solid foundation, the authors acknowledge several open challenges:

  • From Tasks to Goals: Currently, the agent excels at specific tasks (e.g., "price this option"). Moving to open-ended goals (e.g., "earn 2% return") requires the agent to translate vague objectives into concrete, testable hypotheses.
  • Self-Evolution: While the framework preserves research history, the next step is enabling agents to autonomously "learn" from this history—recognizing when a previously successful strategy no longer applies because market regimes have shifted.
  • Semantic Verification: The current system validates numerical claims but does not yet guarantee the correctness of the agent's qualitative financial interpretations or unit labels in generated reports.

Summary updated

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.
BibTeX
@misc{wu2026vibe,
  author = {Haozhe Wu and Zhonghang Li and Chao Huang},
  title = {{Vibe-Trading: Towards Agent-Native Trading}},
  year = {2026},
  eprint = {2611.24845},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.24845},
}
Figure