TaskWeaver is a code-first agent framework designed for data analytics tasks. It uses LLMs to interpret user requests as code snippets and coordinates various plugins as functions to execute tasks in a stateful manner. Unlike typical agent frameworks, it preserves both chat history and code execution history, including in-memory data, enabling expressiveness for complex data structures like high-dimensional tabular data. The framework supports multiple roles including a Planner with vision input and a Recepta role for reasoning, and integrates with AgentOps for observability.