Summary

DeepAgent is a general-purpose reasoning agent designed for autonomous task completion via scalable toolsets. It targets users needing sophisticated AI agents for complex tasks, offering an end-to-end solution for tool discovery, reasoning, and action execution. Its primary benefit is dynamically accessing and utilizing a vast array of tools to tackle diverse challenges, from general API interactions to embodied AI and deep research. The project's paper is available on arXiv 2510.21618.

How It Works

DeepAgent operates as a unified, end-to-end deep reasoning agent, departing from traditional sequential loops. It maintains a global perspective, autonomously discovering and executing tools within a single thought stream. For long-horizon tasks, it employs "Autonomous Memory Folding," compressing interaction history into a structured, brain-inspired memory schema (Episodic, Working, Tool Memory) for strategy reconsideration. Training uses "ToolPO," an end-to-end RL method with an LLM-based tool simulator and fine-grained credit attribution for tool calls.

Quick Start & Requirements

Installation requires Python 3.10, a Conda environment, and pip install -r requirements.txt after cloning into the DeepAgent-main directory. Models must be served, preferably via vLLM, supporting configurations like Qwen3 and QwQ. Extensive API key configuration is needed for ToolBench (RapidAPI), Google Serper, Jina, TMDB