My Trading Agent
Checking…https://www.myfinance5051.com/api

From research prototype to workplace tool

TradingAgents already demonstrates that multi-agent LLM systems can perform specialised, debate-driven financial reasoning. Its current form is oriented toward developers and researchers. This project adds an interactive dashboard, automated visual analytics and executive-style reporting so that non-technical professionals can interpret multi-agent AI output — reducing cognitive load, improving explainability and aligning the system with professional workflows.

Strengths of the current system

Strong modularity
Analyst, researcher, trader and risk roles mirror real organisational specialisation.
Professional reasoning
Bull/bear debate and risk review resemble genuine finance workflows.
Explainability
Intermediate role-specific reasoning can be inspected, unlike single-agent tools.
Flexible engineering base
Multi-provider support, config options, checkpoints and decision memory.

Limitations as a workplace AI tool

Developer-centric interaction
Access is mainly through a CLI and config files.
High cognitive load
Users must read large volumes of agent text to reach the decision.
Limited visual communication
Little built-in graphing or dashboard capability.
Limited end-user automation
No polished batch or scheduled reporting pipeline.
Limited managerial summarisation
Outputs are research-oriented, not executive-oriented.
Audit and traceability gaps
Run provenance matters in finance settings.

What this workbench adds

  1. 1Interface layerBrowser dashboard for ticker, date, provider, models, depth and analyst team.
  2. 2Live execution layerStreams agent status, tool calls and token usage while the graph runs.
  3. 3Visualisation layerPrice trend, agent stance/consensus and recommendation confidence charts.
  4. 4Narration layerExecutive summary, bull/bear points and agent contributions as a manager-facing memo.
  5. 5Audit layerRun metadata: timestamp, provider/model, runtime, token counts and decision trail.

How consensus and confidence are derived

The framework does not emit a native numeric confidence for every agent, so this layer uses documented proxy metrics: the balance of bullish versus bearish arguments, how many role outputs align with the final action, and confidence language extracted from the text. High consensus means most role outputs align; low consensus means major bull/bear conflict. This is an interpretive layer and is labelled as such throughout the interface.

Evaluation plan

Evaluation compares the original CLI baseline against this enhanced system across four dimensions — usability, explainability, automation efficiency and governance/traceability — rather than trading returns. Task-based user testing measures time to complete, perceived difficulty, clarity of outputs and confidence in understanding.

References