Capability · Framework — agents

TaskWeaver

TaskWeaver from Microsoft Research pioneered the 'code interpreter as orchestration' pattern: instead of a ReAct loop of tool calls, the planner writes Python code that composes plugins, and the worker executes it. That yields stateful, data-frame-friendly conversations — ideal for analytics, chart generation, and multi-step data tasks.

Framework facts

Category: agents
Language: Python
License: MIT
Repository: https://github.com/microsoft/TaskWeaver

Install

pip install taskweaver
taskweaver -p ./project   # scaffolds a project dir

Quickstart

from taskweaver.app.app import TaskWeaverApp

app = TaskWeaverApp(app_dir='./project')
session = app.get_session()
resp = session.send_message('plot the sine of 0..2pi')
print(resp.to_dict())

Alternatives

AutoGen — Microsoft's other agent framework
OpenAI Assistants API — code interpreter as-a-service
LangGraph — graph-based alternative
Jupyter AI — notebook-native agent

Frequently asked questions

TaskWeaver vs AutoGen?

AutoGen is conversation-first (agents chat, occasionally use tools). TaskWeaver is code-first (the planner writes Python that executes). TaskWeaver wins on analytics workflows; AutoGen wins on free-form multi-agent patterns.

Do plugins have to be Python?

Yes — TaskWeaver plugins are Python callables registered via decorators. The planner sees their signatures and doc-strings to decide how to call them.

Sources

TaskWeaver — docs — accessed 2026-04-20
TaskWeaver GitHub — accessed 2026-04-20