<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Agent on Peng Tan's AI Blog</title><link>https://68e49aad.hobbytp-github-io.pages.dev/zh/tags/agent/</link><description>一个关注 AI 各领域的专题博客</description><atom:link href="https://68e49aad.hobbytp-github-io.pages.dev/zh/tags/agent/index.xml" rel="self" type="application/rss+xml"/><item><title>OpenEnv: 智能体安全执行环境</title><link>https://68e49aad.hobbytp-github-io.pages.dev/zh/projects/agent/openenv/</link><pubDate>Thu, 04 Dec 2025 22:10:00 +0800</pubDate><guid>https://68e49aad.hobbytp-github-io.pages.dev/zh/projects/agent/openenv/</guid><description>Meta和Hugging Face正在合作推出**OpenEnv Hub**，这是一个共享和开放的社区中心，用于开发、分享和探索兼容OpenEnv的智能体环境，以支持下一波智能体开发浪潮。这些环境可以用于训练和部署，并作为可扩展智能体开发的基石。</description></item><item><title>芮勇博士介绍</title><link>https://68e49aad.hobbytp-github-io.pages.dev/zh/celebrity_insights/yongrui_lenovo/</link><pubDate>Thu, 20 Nov 2025 16:10:00 +0800</pubDate><guid>https://68e49aad.hobbytp-github-io.pages.dev/zh/celebrity_insights/yongrui_lenovo/</guid><description>芮勇博士介绍</description></item><item><title>深入理解AI Agent：芮勇博士的九个核心问题解析</title><link>https://68e49aad.hobbytp-github-io.pages.dev/zh/celebrity_insights/yongrui_agent_nine_question/</link><pubDate>Thu, 20 Nov 2025 16:10:00 +0800</pubDate><guid>https://68e49aad.hobbytp-github-io.pages.dev/zh/celebrity_insights/yongrui_agent_nine_question/</guid><description>联想芮勇博士对AI Agent的未来提出了九个深刻而关键的问题</description></item><item><title>Agent训练新范式：Agent Learning via Early Experience</title><link>https://68e49aad.hobbytp-github-io.pages.dev/zh/meta/agent_learn_by_early_experience/</link><pubDate>Tue, 14 Oct 2025 18:00:00 +0800</pubDate><guid>https://68e49aad.hobbytp-github-io.pages.dev/zh/meta/agent_learn_by_early_experience/</guid><description>传统AI训练像是把人类所有的知识都强行灌输AI，而Meta的最新论文《Agent Learning via Early Experience》为我们展示了一条训练AI智能体的新路径: 可扩展、无需奖励的实用范式，通过将智能体自身的行为和结果转化为强大的监督信号，显著提升了AI的性能、数据效率和泛化能力。</description></item><item><title>AI Context Engineering领域若干重要论文</title><link>https://68e49aad.hobbytp-github-io.pages.dev/zh/papers/context_engineering_collection/</link><pubDate>Sun, 12 Oct 2025 20:10:00 +0800</pubDate><guid>https://68e49aad.hobbytp-github-io.pages.dev/zh/papers/context_engineering_collection/</guid><description>这里将收集Context Engineering相关的重要文献，具体解读将在其他博客展开。</description></item><item><title>Agent 相关课程收集</title><link>https://68e49aad.hobbytp-github-io.pages.dev/zh/courses/mas_courses/</link><pubDate>Thu, 28 Aug 2025 20:10:00 +0800</pubDate><guid>https://68e49aad.hobbytp-github-io.pages.dev/zh/courses/mas_courses/</guid><description>搜集网络上来自各知名机构的Agent相关课程。</description></item><item><title>Agent Lightning</title><link>https://68e49aad.hobbytp-github-io.pages.dev/zh/mas/agent_lightning/</link><pubDate>Wed, 27 Aug 2025 20:10:00 +0800</pubDate><guid>https://68e49aad.hobbytp-github-io.pages.dev/zh/mas/agent_lightning/</guid><description>&lt;h2 id="介绍">介绍&lt;/h2>
&lt;p>微软开源的 &lt;strong>Agent Lightning&lt;/strong> 项目，它的核心价值在于为开发者和研究者提供了一个强大的工具，用于&lt;strong>训练和优化 AI Agent（智能代理）&lt;/strong>，特别是&lt;strong>几乎不需要修改现有 Agent 代码&lt;/strong>就能实现显著的性能提升。&lt;/p>
&lt;p>这个项目有以下重要作用：&lt;/p>
&lt;ol>
&lt;li>
&lt;p>&lt;strong>零代码/低代码训练 AI Agent (核心价值)：&lt;/strong>&lt;/p>
&lt;ul>
&lt;li>&lt;strong>最大亮点：&lt;/strong> 它允许你使用&lt;strong>强化学习(Reinforcement Learning, RL)&lt;/strong> 等高级优化算法来训练你现有的 AI Agent，而&lt;strong>几乎不需要修改你的 Agent 业务逻辑代码&lt;/strong>。这意味着你可以保留你用 LangChain, AutoGen, CrewAI, OpenAI SDK 等框架（甚至裸 Python）编写的 Agent 逻辑，然后让 Agent Lightning 负责优化它的决策过程。&lt;/li>
&lt;li>&lt;strong>解决痛点：&lt;/strong> 传统上，将 RL 等技术应用到现有 Agent 框架中需要大量的工程改造和集成工作。Agent Lightning 极大地简化了这个过程。&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>
&lt;p>&lt;strong>强大的优化能力：&lt;/strong>&lt;/p>
&lt;ul>
&lt;li>&lt;strong>算法支持：&lt;/strong> 内置支持&lt;strong>强化学习(VERL)&lt;/strong> 作为核心优化算法，并明确提到支持&lt;strong>自动提示优化(Automatic Prompt Optimization, APO)&lt;/strong>。未来很可能扩展更多算法。&lt;/li>
&lt;li>&lt;strong>提升性能：&lt;/strong> 通过优化，Agent 在执行任务（如 SQL 生成与修正、工具调用、复杂决策）时的准确性、效率和可靠性可以得到显著提升。&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>
&lt;p>&lt;strong>广泛的兼容性和灵活性：&lt;/strong>&lt;/p>
&lt;ul>
&lt;li>&lt;strong>框架无关：&lt;/strong> 明确支持所有主流 Agent 框架（LangChain, OpenAI Agent SDK, AutoGen, CrewAI）以及纯 Python 实现的 Agent。你可以“即插即用”。&lt;/li>
&lt;li>&lt;strong>多 Agent 系统优化：&lt;/strong> 可以在包含多个 Agent 的复杂系统中，&lt;strong>选择性地优化其中一个或几个特定的 Agent&lt;/strong>，而不是整个系统，提供了更精细的控制。&lt;/li>
&lt;/ul>
&lt;/li>
&lt;li>
&lt;p>&lt;strong>提供训练基础设施：&lt;/strong>&lt;/p></description></item><item><title>Claude Code 介绍以及开源生态</title><link>https://68e49aad.hobbytp-github-io.pages.dev/zh/claude/claude_code/</link><pubDate>Sun, 20 Jul 2025 23:10:00 +0800</pubDate><guid>https://68e49aad.hobbytp-github-io.pages.dev/zh/claude/claude_code/</guid><description>Claude Code 是 Claude 的命令行工具，用于代理编码，提供灵活的、可定制的、可脚本化的和安全的编程方式。</description></item><item><title>AI for Research Survey</title><link>https://68e49aad.hobbytp-github-io.pages.dev/zh/papers/ai_for_research_survey/</link><pubDate>Sun, 06 Jul 2025 23:03:00 +0800</pubDate><guid>https://68e49aad.hobbytp-github-io.pages.dev/zh/papers/ai_for_research_survey/</guid><description>AI for Research Survey. 这篇综述文章概述了人工智能在科学研究（AI4Research）中的应用，强调了大型语言模型（LLMs）在促进科学发现各个阶段的潜力。它建立了一个系统的分类体系，将AI4Research分解为五个主要任务：科学理解、学术调查、科学发现、学术写作和学术同行评审。文章详细介绍了AI在这些领域中的具体应用，例如辅助思想挖掘、实验设计、数据分析、论文撰写和同行评审。此外，该综述还识别了未来的研究方向，包括跨学科AI模型、伦理考量以及多模态和多语言集成，旨在为研究界提供资源并激发AI驱动的科学创新。</description></item><item><title>Agent经济：红杉资本2025 AI峰会释放的超级信号</title><link>https://68e49aad.hobbytp-github-io.pages.dev/zh/my_insights/ai_ascent_2025/</link><pubDate>Tue, 13 May 2025 15:10:00 +0800</pubDate><guid>https://68e49aad.hobbytp-github-io.pages.dev/zh/my_insights/ai_ascent_2025/</guid><description>Agent经济：红杉资本2025 AI峰会释放的超级信号</description></item><item><title>模型上下文协议（MCP）深度解析：Agent互操作性的新纪元</title><link>https://68e49aad.hobbytp-github-io.pages.dev/zh/claude/mcp_analysis/</link><pubDate>Tue, 29 Apr 2025 20:10:00 +0800</pubDate><guid>https://68e49aad.hobbytp-github-io.pages.dev/zh/claude/mcp_analysis/</guid><description>本文介绍了模型上下文协议（MCP），并对其技术原理、主要贡献、当前优劣、生态系统现状，并与Google A2A等相关技术进行比较，展望其未来发展趋势。</description></item><item><title>模型上下文协议（MCP）深度解析：Agent互操作性的新纪元</title><link>https://68e49aad.hobbytp-github-io.pages.dev/zh/claude/mcp_usecases/</link><pubDate>Tue, 29 Apr 2025 20:10:00 +0800</pubDate><guid>https://68e49aad.hobbytp-github-io.pages.dev/zh/claude/mcp_usecases/</guid><description>本文介绍了模型上下文协议（MCP），并对其技术原理、主要贡献、当前优劣、生态系统现状，并与Google A2A等相关技术进行比较，展望其未来发展趋势。</description></item><item><title>Deep Research 深度研究</title><link>https://68e49aad.hobbytp-github-io.pages.dev/zh/draft/deepresearch/</link><pubDate>Sun, 13 Apr 2025 15:10:00 +0800</pubDate><guid>https://68e49aad.hobbytp-github-io.pages.dev/zh/draft/deepresearch/</guid><description>Deep Research 深度研究</description></item><item><title>Google AI协同科学家系统</title><link>https://68e49aad.hobbytp-github-io.pages.dev/zh/google_ai_co-scientist/</link><pubDate>Thu, 20 Feb 2025 22:40:00 +0800</pubDate><guid>https://68e49aad.hobbytp-github-io.pages.dev/zh/google_ai_co-scientist/</guid><description>本文介绍了Google开发的AI协同科学家系统（AI co-scientist），并对其技术原理、主要贡献、论文方法、评估结果和局限性进行了详细解读。</description></item></channel></rss>