QitOS Documentation

The torch-flavor framework for agent researchers building reproducible LLM agents on one AgentModule + Engine kernel.

QitOS is a PyTorch-style framework for agent researchers. It lets you prototype methods, run benchmarks, and inspect long-horizon trajectories on a single AgentModule + Engine kernel, with built-in observability via qita (the trace viewer that ships with QitOS).

Two authoring paths

QitOS offers two top-level ways to build agents:

• Research-first: handwrite the prompt, parser, protocol (the output format contract between model and parser), transport (the client that talks to the model API), and tool surface so you can study and optimize the kernel directly.
• Preset-first: start from family presets (pre-configured model-family defaults) and preset tool builders when you want a fast baseline or quick multi-family switching.

Both paths are first-class citizens in the framework.

Quick Start

Run your first agent in under 2 minutes

Build Your First Agent

Step-by-step walkthrough to build a working agent

Tutorial Track

Learn QitOS through four course-style agent builds

Core Concepts

Understand AgentModule, Engine, State, and Tools

API Reference

Full reference for the public QitOS API

Research-first path

Handwrite prompt, parser, protocol, and tool surface on top of the kernel.

Preset-first path

Use family presets to switch one example across multiple model families.

Get up and running

1. Install QitOS

   pip install "qitos[models]"
   
2. Set your provider config

   export OPENAI_API_KEY="sk-..."
   export OPENAI_BASE_URL="https://api.siliconflow.cn/v1/"
   export QITOS_MODEL="Qwen/Qwen3-8B"
   
3. Run the minimal coding agent

   qit demo minimal
   
4. Inspect your run with qita

   qita board --logdir runs
   

   Open http://127.0.0.1:8765 to view trace, replay, and export your run.

What QitOS gives you

Agent Patterns

ReAct, PlanAct, Tree-of-Thought, and Reflexion out of the box

Course-style tutorials

ReAct, PlanAct, Claude Code-style, and security audit lessons

Observability

Built-in tracing with qita board, replay, and HTML export

Benchmarks

GAIA, Tau-Bench, and CyBench adapters on the same kernel

Kit Toolkit

Parsers, memory adapters, toolsets, critics, and planners

Featured designs

Single kernel

One AgentModule + Engine runtime across examples, benchmarks, and production-style agents

Protocol-aware parsers

Prompt format and parser choice are paired intentionally, not left as implicit glue code

Long-running context control

History policy, summary history, compact history, and memory adapters for sustained runs

Traceability with qita

Board, replay, and export built on top of structured trace artifacts