Agent

This section aggregates courses and technical frameworks related to LLM agents.

OpenHands (formerly OpenDevin)

Links:
- GitHub Repository
- Paper arXiv:2407.16741
Overview:
OpenHands is an open-source Agent platform capable of executing code / shell / web / API tasks, suited for developers' daily automation and toolchain integration.

Links:
- Official Introduction
- Zhihu Article
Overview:
A reinforcement learning (RL)-driven research Agent that can self-plan, perform multi-round retrieval and reasoning, for complex tasks such as technical research and report writing.

Links:
- Introducing Deep Research — OpenAI
- ChatGPT Deep Research — Wikipedia
Overview:
Deep Research is an Agent mode launched by OpenAI in 2025, supporting multi-step internet research tasks with the ability to browse web pages, PDFs, and image materials to synthesize reports. Suited for professional research in fields such as science, policy, and engineering.

Links:
- arXiv:2409.13731 (KAG)
Overview:
KAG is an LLM enhancement framework for professional domains, designed to combine knowledge graphs + vector retrieval to address the limitations of RAG in specialized scenarios. It places particular emphasis on the integration of knowledge logic (such as numerical values, temporal information, and expert rules) to deliver more reliable performance on question-answering and reasoning tasks.

Links:
- Deep Research Agents: A Systematic Examination and Roadmap (arXiv)
- DeepResearcher: Scaling Deep Research via Reinforcement Learning (arXiv:2504.03160)
Overview:
These works study the architectural components, retrieval methods, tool usage, and planning processes for Agents performing deep research in real-world environments.
- The DR Survey provides a taxonomy of Agent systems, key challenges, and future directions.
- DeepResearcher trains Agents using RL + web interaction, outperforming traditional retrieval- or prompt-based methods on open-domain research tasks.

这篇文章有帮助吗？