An AI research intern that actually reads balance sheets

What it does

Dexter is an autonomous agent that decomposes complex financial questions into step-by-step research tasks. It executes those tasks by calling tools for live income statements, balance sheets, and cash flow data, then validates its own results and refines them until it reaches a confident answer. Safety rails like loop detection and step limits keep the agent from spiraling on a bad query.

The interesting bit

The project treats agent reasoning as an audit trail. Every query writes a JSONL scratchpad to .dexter/scratchpad/ that records the original question, each tool call with arguments and raw results, and the LLM’s interpretation—so you can debug exactly how it reached a conclusion. It also ships with an evaluation suite that scores the agent against a dataset of financial questions using LangSmith and an LLM-as-judge approach.

Key highlights

• Automatically breaks down complex financial queries into structured research steps
• Pulls real-time financial data via the Financial Datasets API
• Supports multiple LLM backends: OpenAI, Anthropic, Google, xAI, OpenRouter, and local Ollama
• Built-in loop detection and step limits prevent runaway execution
• Ships with an evaluation suite that uses LangSmith and an LLM-as-judge for scoring

Caveats

• Requires multiple API keys (OpenAI, Financial Datasets, and optionally Exa) to function
• Outputs may be incorrect, incomplete, or out of date; the README explicitly disclaims real trading or investment use
• Built specifically for the Bun runtime, not Node.js

Verdict

Worth a look for developers studying autonomous agents with structured tool use and self-correction. Avoid if you need a free, offline, or legally-sound investment advisor.