TaskWeaver — Code-First Data Analytics Agent

What it is

TaskWeaver is a code-first agent framework by Microsoft designed for data analytics tasks. It takes natural language requests, decomposes them into a plan, generates Python code for each step, executes the code in a stateful session, and returns results with DataFrames and visualizations.

TaskWeaver targets data analysts and engineers who want an AI agent that writes and runs analytics code rather than just answering questions in text.

How it saves time or tokens

TaskWeaver generates and executes code in a persistent session. Variables, DataFrames, and computation results persist across steps. This means the agent builds on previous work instead of starting fresh each time, reducing redundant computation and token usage.

The plugin system lets you extend TaskWeaver with domain-specific functions (database connectors, API calls, custom transformations) that the agent can invoke by name.

How to use

1. Clone the TaskWeaver repository
2. Configure your LLM provider in the config file
3. Add plugins for your data sources (SQL connectors, file readers, APIs)
4. Run TaskWeaver and describe your analytics task in natural language

Example

# TaskWeaver plugin: sql_query.py
from taskweaver.plugin import Plugin, register_plugin

@register_plugin
class SQLQuery(Plugin):
    def __call__(self, query: str) -> dict:
        '''Execute a SQL query and return results as a DataFrame.'''
        import pandas as pd
        import sqlalchemy
        engine = sqlalchemy.create_engine(self.config['db_url'])
        df = pd.read_sql(query, engine)
        return {'data': df, 'row_count': len(df)}

# User interaction:
# > Load sales data from the database and show monthly revenue trends
# TaskWeaver: Planning...
# Step 1: Query sales table using sql_query plugin
# Step 2: Group by month and calculate revenue
# Step 3: Generate a line chart

Related on TokRepo

• Database tools -- Data access and query tools
• AI agent tools -- Agent frameworks

Common pitfalls

• TaskWeaver executes arbitrary Python code; run it in a sandboxed environment, never on production machines with sensitive data access
• The planner sometimes generates overly complex plans for simple tasks; provide clear, specific instructions
• Plugin configuration requires matching the exact function signature that TaskWeaver expects; follow the plugin template