ANative Lab

Building a Self-Evolving Ecosystem of AI Agents.

[ANative Lab @ANativeLab]

📣 EvoAgentX has a new home: ANative Lab.

You'll notice a new name & avatar on this account — same project, same mission, still open-source & MIT. EvoAgentX is now the flagship open-source framework of ANative Lab, a research lab for AI-native & agent-native intelligence, advancing self-evolving agents and autonomous scientific discovery.

ANative Lab is more than one project:
🔹 EvoAgentX — Build. Evaluate. Evolve. Open-source self-evolving agent workflows.
🔹 EvoScientist — your self-evolving AI scientist: a multi-agent system for end-to-end scientific discovery, by @_xizhang.

The team:
• @mengzaiqiao — Founder & Lab Director
• @TedSiwei — Co-Lead
• @JinyuanF — Co-Lead
• @_xizhang — Chief Scientist
• @Zhaohan_Meng — Research Scientist

🌐 https://t.co/6lLVgQVCtD
💻 https://t.co/Tp97wWbIVA

[ANative Lab @ANativeLab]

Missed the talk? Check out the meeting record here: https://t.co/TitEW9IvS1

[ANative Lab @ANativeLab]

🎮 Can an AI agent keep getting better while it's still running?

That's the question behind our next EvoAgentX Talk.

We're thrilled to host Seth Karten (PhD @ Princeton, researcher @ Prime Intellect) on Continual Harness: Online
Adaptation for Self-Improving Foundation Agents.

Instead of fixed prompts and static workflows, Continual Harness lets agents adapt online — accumulating experience,
refining their behavior, and reusing what they learn across long-horizon tasks. Seth will show it live through Gemini
Plays Pokémon (where harness-level improvements dramatically boost game progress), extend it to ARC-AGI-3, and
generalize it to a coding agent that learns from its own past attempts.

🗓️ Thu, Aug 6, 2026 · 10:00 AM (GMT+8) / Wed, Aug 5 · 7:00 PM PDT
💻 Microsoft Teams

Join us 👇 details in the comments.

#AIAgents #FoundationModels #LLM #AgenticAI #EvoAgentX

[ANative Lab @ANativeLab]

Great work!

[Seth Karten @sethkarten]

Prime Agent, our new general CLI agent, is out. We show a huge narrative violation where models have the potential to outperform their native harnesses across a variety of evals

Huge shoutout to @a1zhang @omouamoua @kevinjosethomas for working on this together

[ANative Lab @ANativeLab]

@FReza1984 Yes — if the adaptation is verification-gated: the agent reinforces only what provably passes tests, keeps failures as negative examples, and since it lives in the harness rather than the weights, it stays auditable and reversible.

[ANative Lab @ANativeLab]

🎮 Can an AI agent keep getting better while it's still running?

That's the question behind our next EvoAgentX Talk.

We're thrilled to host Seth Karten (PhD @ Princeton, researcher @ Prime Intellect) on Continual Harness: Online
Adaptation for Self-Improving Foundation Agents.

Instead of fixed prompts and static workflows, Continual Harness lets agents adapt online — accumulating experience,
refining their behavior, and reusing what they learn across long-horizon tasks. Seth will show it live through Gemini
Plays Pokémon (where harness-level improvements dramatically boost game progress), extend it to ARC-AGI-3, and
generalize it to a coding agent that learns from its own past attempts.

🗓️ Thu, Aug 6, 2026 · 10:00 AM (GMT+8) / Wed, Aug 5 · 7:00 PM PDT
💻 Microsoft Teams

Join us 👇 details in the comments.

#AIAgents #FoundationModels #LLM #AgenticAI #EvoAgentX

[ANative Lab @ANativeLab]

📌 Join details:

💻 Microsoft Teams
🗓️ Thu, Aug 6, 2026 · 10:00 AM (GMT+8) / Wed, Aug 5 · 7:00 PM PDT

Meeting ID: 353 021 243 807 410
Passcode: 7Ds2vD98

See you there! Feel free to share with anyone into agents & self-improving AI. 🚀

[ANative Lab @ANativeLab]

EAX 讲座预告｜从固定记忆操作到可演化记忆技能：MemSkill 驱动智能体记忆自主进化

讲座核心内容
长时序交互场景下，智能体长效能力的核心瓶颈不再只是经验存储，而是缺少自主优化记忆方式的能力。传统智能体记忆系统全部依赖人工预设、一成不变的记忆操作，无法适配动态任务与多变交互场景。本期 EvoAgentX Talk 特邀南���理工大学博士生张昊臻，独家分享原创 MemSkill 记忆自演化框架，解锁智能体自主学习记忆策略的全新技术路径。

嘉宾简介
张昊臻@haozhen_ntu，南洋理工大学（NTU）计算机科学在读博士生。研究方向聚焦长程智能体系统、经验结构化复用与智能体自我演化，原创提出 MemSkill 智能体记忆自演化框架，实现记忆技能自主搭建与迭代升级。 先后于伊利诺伊大学厄巴纳 - 香槟分校、香港科技大学、微软亚洲研究院开展科研实习与合作研究，多项研究成果收录 ICML、NeurIPS、ACL 等人工智能领域顶会，深耕长效智能体记忆前沿赛道。

分享要点
长效智能体想要真正实现持续进化，除积累交互经验外，更要学会自主迭代 “记忆方式”，现有记忆方案存在固有局限：
传统智能体记忆模块依靠人工固化记忆操作，灵活性不足，面对差异化任务、动态交互环境适配能力极差；
本次主讲 MemSkill 创新框架，把僵硬固定的记忆操作改造为可训练、可组合、可持续演化的记忆技能，智能体可在交互过程中自动筛选、优化记忆策略；
框架支持从历史任务、失败交互样本中提炼全新记忆策略，让智能体动态适配当下环境与任务需求；
区别于传统只侧重经验存储检索的记忆架构，MemSkill 聚焦记忆能力本身的持续优化，推动智能体从被动存放过往信息，升级为主动学习记忆方法，完成记忆层面的完整自我演化闭环。

活动信息
📅 讲座时间：2026 年 06 月 28 日 16:30-17:30（GMT+8 中国标准时间）
✨ 主讲人主页：https://t.co/azbaiw8fqI
🔍 分享主题：从固定记忆操作到可演化记忆技能：MemSkill 与智能体记忆自演化
💻 会议：腾讯会议 | 会议号：864-827-250
🔗 入会链接：https://t.co/5eDKlRHutb

如果你深耕长程智能体、智能体记忆架构、持续自演化 Agent、记忆优化策略等研究方向，本场分享将带来记忆层自主进化的创新解法，值得深度聆听！ 6 月 28 日，和张昊臻一同探索智能体记忆自主迭代的前沿方案～
#AI #Agent #LLM #MemSkill

[ANative Lab @ANativeLab]

EAX 讲座预告｜COS-PLAY：长程任务下 LLM 智能体的协同演化新范式

当下大模型智能体在长程交互场景中，普遍存在技能难以沉淀、复用与迭代的痛点，极大限制了智能体长期决策能力的提升。EvoAgentX Talk 全新专场来袭，特邀马里兰大学帕克分校博士吴曦旸，带来COS-PLAY协同演化智能体框架深度分享，破解长程任务中智能体技能积累难题。

讲座核心内容
本次分享针对 LLM 智能体长期决策里，技能无法稳定挖掘、留存和复用的行业痛点，全面解读 COS-PLAY 框架的设计思路与技术细节：
• 拆解双智能体协作机制：决策智能体负责检索技能、下达行动指令，技能库智能体依托无标注交互轨迹，自动完成技能抽取、精炼与动态更新，并划定技能使用约束；
• 展示多场景实验成果：依托六大游戏环境开展验证，该框架不仅在单人长程任务中效果提升显著，在多人社交推理场景也具备亮眼表现；
• 探索技术落地价值，为打造拥有长效技能积累能力的大模型智能体，提供全新的研究思路与实现方案。

嘉宾简介
吴曦旸https://t.co/UygwqF6ZMa，马里兰大学帕克分校��子与计算机工程系五年级博士生，师从 Dinesh Manocha 教授。主要研究方向为多模态大模型与复杂智能体系统，深耕模型幻觉检测、物理推理、多模态长上下文推理等领域。 现阶段他聚焦长程智能体决策、自演化智能体、强化学习等方向，致力于研究如何让智能体依托记忆沉淀经验、迭代技能，在动态不确定环境中持续优化决策能力，推动大模型智能体从被动执行，进化为可自主规划、终身学习、自我优化的高阶智能系统。

活动信息
📅 讲座时间：2026 年 06 月 21 日 20:00-21:00（GMT+8 中国标准时间）
✨ 主讲人主页：https://t.co/aIFG315iRN
🔍 分享主题：COS-PLAY：面向长程任务的 LLM 决策智能体与技能库智能体协同演化
💻 会议：腾讯会议 | 会议号：514-245-531
🔗 入会链接：https://t.co/jjXH4lJO39

如果你关注长程智能体、智能体协同演化、技能库构建、大模型持续学习等前��方向，本场讲座不容错过。6 月 21 日，一同探索长程任务智能体的进化之路！

#AI #AIAgent #LLM #EvoAgentX #HowOne

[ANative Lab @ANativeLab]

EAX讲座预告｜自演化智能体 SkillOS 框架深度分享

大模型智能体迈入长周期落地新阶段，静态技能架构已然瓶颈。EvoAgentX Talk 重磅特邀 UIUC 顶尖在读博士欧阳思如，围绕原创自演化智能体框架 SkillOS 展开独家学术分享，拆解智能体自主习得、技能迭代、元技能涌现核心技术，聚焦当下 Agent 领域前沿痛点与创新解法。

嘉宾简介
欧阳思如，伊利诺伊大学香槟分校 CS 四年级博士生，师从数据库与 AI 领域泰斗韩家炜教授。主攻智能体系统、记忆架构与持续学习方向，先后于微软雷德蒙德、腾讯西雅图、Google、NVIDIA 四大国际科技巨头实习深耕；科研成果落地多篇顶会论文，收录 ICML、ICLR、ACL、EMNLP、KDD 等人工智能、NLP、数据挖掘��域顶刊顶会，学术与工业落地经验兼备。@Siru_Ouyang

分享要点
随着 LLM 智能体大规模投入长时序、流式动态环境，智能体的上限不再局限单次任务执行效果，经验沉淀与终身持续学习成为关��。现有主流技能型 Agent 普遍把技能固化为静态知识库，无法跟随环境与任务动态更新优化。本场报告主讲自研SkillOS 自演化框架，开创性把技能管控拆分为独立学习目标，解耦任务执行模块与可训练技能管理器，借助经验强化学习自主完成技能入库、修正、淘汰全流程。实测数据对比传统存储方案优势突出，在提升技能调用准确率的同时，驱动智能体在不间断交互中自然生成高阶元技能，为打造可自主进化、长效迭代的通用智能体开辟全新技术范式。

活动信息
📅 讲座时间：06月 07 日 21:00-22:00
✨ 主讲人主页：https://t.co/OotTEwC6XN
💻 会议：腾讯会议 | 会议号：451-239-675
🔗 入会链接：https://t.co/7G8uQJcJZG

#AI #AIAgent #SkillOS #EvoAgentX

[ANative Lab @ANativeLab]

Two powerful upgrades on HowOne to make AI app building effortless ✨

• Curated Templates: Explore our handpicked high-quality templates. Jump right in and start instantly, no need to build from scratch.

• Improved Onboarding: Brand new guided steps for app creation. Follow the clear workflow and build your AI app with total confidence.

#HowOne #AI #AIApps #AIAgent #NoCode #LowCode #AppBuilder #BuildInPublic #GenerativeAI #AIWorkflow

[ANative Lab @ANativeLab]

HowOne just got a major upgrade 🚀

We’ve refreshed the brand-new homepage with a brand new visual design.Cleaner layout and clearer structure let you quickly figure out what you can build with HowOne in no time.

[ANative Lab @ANativeLab]

We have reached 3k stars!

[ANative Lab @ANativeLab]

EAX讲座预告｜语言智能体的下半场：从“能做事”到“会决策”

当语言智能体度过“能完成任务”的启蒙阶段，下一个核心突破点在哪里？答案，或许藏在“学会做正确决策”的进阶之路中。
本周EvoAgentX Talk，我们荣幸邀请到爱丁堡大学博士后Hongru Wang，带来《语言智能体的下半场》主题分享，从Agent理论（ToA）出发，带我们解锁语言智能体的进阶密码，展望终身学习型智能体的未来图景。

讲座核心内容
本次分享将聚焦语言智能体的“下半场”转型，系统拆解从“能做事”到“会决策”的核心逻辑与研究路径：
明确语言智能体的发展分界：上半场聚焦“打造有能力的智能体”（完成既定任务），下半场核心是“打造有智慧的智能体”（做出优质决策）；
深度解析Agent理论（ToA）：将智能体的内部推理与外部行动统一为两种认知等价的工具，建模参数空间中的内部世界与外部物理世界，打破“工具增强文本生成器”的固有认知；
覆盖智能体优化全场景：从智能体监督微调（agentic SFT）、智能体强化学习（agentic RL）到智能体提示学习（agentic prompting），提供系统的智能体行为优化框架；
展望终身语言智能体愿景：构建可持续学习、与用户及环境保持对齐、能自主演化内部结构（如记忆）的智能系统，适配复杂真���世界场景。

主讲人简介
Hongru Wang
爱丁堡大学博士后研究员（Research Associate），任职于EdinburghNLP与EdinburghAI，紧密合作导师为Amos Storkey教授与Jeff Z. Pan教授。
他于香港中文大学获得博士学位，师从ACL会士Kam-Fai Wong教授；博士期间曾在爱丁堡大学EdinburghNLP、BlenderLab及伊利诺伊大学厄巴纳-香槟分校开展研究工作，与Jeff Z. Pan、Heng Ji、Irwin King、Mengdi Wang等多位知名教授深度合作。
作为NLP学术交流平台（NICE）的联合创始人及组织者，该平台已汇聚15万+AI与NLP领域爱好者，成为行业重要的学术分享阵地。其研究核心聚焦于Agent理论（ToA，相关成果将发表于ICML 2026），长期目标是破解语言智能体“安全性、个性化、自主性”的不可能三角，实现智能体从内外部交互中自主学习。

活动信息
📅 时间：2026年05月10日（周日）16:30–17:30（GMT+8，中国标准时间）
🔍 主题：EvoAgentX Talk: 语言智能体的下半场（Second Half of Languages Agent）
💻 会议：腾讯会议 | 会议号：587-622-848
🔗 入会链接：https://t.co/m1qRbBpPsr
如果你关注语言智能体、Agent理论、强化学习、终身学习智能体等前沿方向，这场讲座将为你提供全新的系统视角，带你看清语言智能体的未来发展方向！
5月10日，与Hongru Wang一起，共探语言智能体的下半场征程～

[ANative Lab @ANativeLab]

EAX 讲座预告｜通过状态反思进行学习的智能体框架

大语言模型智能体正从一次性任务执行器，迈向可长期部署、持续进化的通用智能系统。如何让 Agent 在不更新模型参数的前提下，从交互中自主学习、持续变强？这正是当前 Agent 领域的核心难题。

本周 EvoAgentX Talk，我们特别邀请到伦敦大学学院（UCL）博士周辉池，带来《通过状态反思进行学习的智能体框架》主题分享，揭秘基于记忆与反思的 Agent 长效学习新范式。

讲座核心内容
本次分享将围绕无参数更新的持续学习展开，系统讲解如何让智能体在长期交互中自主进化：
• 提出状态反思决策过程（Stateful Reflective Decision Process），将学习建模为带演化记忆的序贯决策系统，从理论层面奠定长效学习基础。
• 详解Memento 智能体记忆微调框架：无需微调基座 LLM，仅优化记忆架构，即可大幅提升复杂长期任务的决策能力与综合性能。
• 重磅介绍Memento‑Skills 系统：把能力外部化为可演化技能库，支持 Agent 自主设计 Agent，具备从 CLI 到 GUI 的完整工程落地能力，适配真实世界长期任务代理。

嘉宾简介
周辉池伦敦大学学院（UCL）博士生，师从汪军教授研究方向：大语言模型智能体（LLM Agents）、智能体记忆（Agentic Memory）。作为核心研究者，他主导提出Memento 智能体记忆微调框架，突破传统模型优化瓶颈，实现无需基座微调的 Agent 性能跃升；相关成果已发表于ICLR、ICML、ACL、KDD、AAAI、IEEE TNNLS等国际顶会与顶刊，是 Agent 记忆与长效学习领域的新锐研究者。

活动信息
📅 时间：2026 年 04 月 26 日（周日）16:30–17:30（GMT+8，中国标准时间）
🔍 主题：EvoAgentX Talk: 通过状态反思进行学习的智能体框架
💻 会议：腾讯会议 | 会议号：938-474-365
🔗 入会链接：https://t.co/RM041usXk9

如果你深耕LLM Agent、智能体记忆、长效学习、自主进化 Agent，这场讲座将为你打开无参数更新的 Agent 进化新路径，不容错过！
4 月 26 日，一起解锁智能体的反思式学习密码～

[ANative Lab @ANativeLab]

What happens when AI scientists can evolve themselves?

@_xizhang will share EvoScientist — a system where multi-agent AI collaborates to solve real research problems and continuously improves itself.

If you're building AI agents, doing research, or exploring self-evolving systems — this is worth your time.

Join us and see how “vibe research” could become reality.

📅 Apr 5, 16:30–17:30 (CN)
🔗 Meeting ID: 470-317-001

[ANative Lab @ANativeLab]

For the past two years, AI founders have been told a comforting story:
Build a good vertical app, pick a niche, wrap the model in a nice UI, charge a subscription.If you execute well, you’ll be fine.

That story is dying.

Not because models got smarter. Because Skills changes what “a product” is.

When Anthropic introduced Claude Skills—and when Agent Skills quickly moved toward an open standard—most people heard “another developer feature.” A new way to package prompts and scripts. Cute.
But if you zoom out, Skills is not a feature. It’s a structural shift:
Skills turns capabilities into installable modules.

And once that happens, a huge percentage of AI startups stop being businesses… and start being plugins that platforms can bundle.
This piece is not a tutorial designed to make you feel optimistic.

It’s a map of the new power dynamics—especially if you’re building an AI product and hoping the platform layer won’t swallow you.

[HowOne @HowOneAI]

https://t.co/haCAagt56X

[Chris Laub @ChrisLaubAI]

🚨 RIP prompt engineering.

This new Stanford paper just made it irrelevant with a single technique.

It's called Verbalized Sampling and it proves aligned AI models aren't broken we've just been prompting them wrong this whole time.

Here's the problem: Post-training alignment causes mode collapse. Ask ChatGPT "tell me a joke about coffee" 5 times and you'll get the SAME joke. Every. Single. Time.

Everyone blamed the algorithms. Turns out, it's deeper than that.

The real culprit? 'Typicality bias' in human preference data. Annotators systematically favor familiar, conventional responses. This bias gets baked into reward models, and aligned models collapse to the most "typical" output.

The math is brutal: when you have multiple valid answers (like creative writing), typicality becomes the tie-breaker. The model picks the safest, most stereotypical response every time.

But here's the kicker: the diversity is still there. It's just trapped.

Introducing "Verbalized Sampling."

Instead of asking "Tell me a joke," you ask: "Generate 5 jokes with their probabilities."

That's it. No retraining. No fine-tuning. Just a different prompt.

The results are insane:

- 1.6-2.1× diversity increase on creative writing
- 66.8% recovery of base model diversity
- Zero loss in factual accuracy or safety

Why does this work? Different prompts collapse to different modes.

When you ask for ONE response, you get the mode joke. When you ask for a DISTRIBUTION, you get the actual diverse distribution the model learned during pretraining.

They tested it everywhere:

✓ Creative writing (poems, stories, jokes)
✓ Dialogue simulation
✓ Open-ended QA
✓ Synthetic data generation

And here's the emergent trend: "larger models benefit MORE from this."

GPT-4 gains 2× the diversity improvement compared to GPT-4-mini.

The bigger the model, the more trapped diversity it has.

This flips everything we thought about alignment. Mode collapse isn't permanent damage it's a prompting problem.

The diversity was never lost. We just forgot how to access it.

100% training-free. Works on ANY aligned model. Available now.

Read the paper: arxiv. org/abs/2510.01171

The AI diversity bottleneck just got solved with 8 words.

[HowOne @HowOneAI]

https://t.co/haCAagt56X

[ANative Lab @ANativeLab]

https://t.co/VzUMxMIGgF