EvoAgentX

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：lnkd.in/gWWCuSj2 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～

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

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.

🚨 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.

🚀 EvoAgentX Transformed LLMs into Autonomous Scientists — And They Just Wrote 6 Research Papers for ICAIS 2025! What happens when you stop treating LLMs as text generators — and start using them as autonomous co-authors? We found out! With EvoAgentX, we built the full-stack agentic systems that wrote, revised, and prepared six technical papers—all submitted to ICAIS 2025, with one of them even receiving a like and a signed “Nice Job” from James J. Heckman, the 2000 Nobel Prize winner in Economics. 💡 What did we build? 🚀 Using two end-to-end AI systems built with EvoAgentX, we automated the full paper development pipeline — from ideation to LaTeX-ready submissions. 1️⃣ Idea-to-Proposal Generator • Multi-agent debate engine to generate & refine ideas • AI-generated researcher personas with domain knowledge • Structured high-quality research proposals 2️⃣ Experiment Execution System • Automatically splits a research plan into stages: initialization, training, tuning, ablations. • Generates summaries, tables, and plots at each stage using API-driven evaluation. • Produces LaTeX-compatible sections with visualizations for seamless paper assembly. 💡 What EvoAgentX created? 📄 We submitted 6 papers to ICAIS 2025 All generated with EvoAgentX + targeted human verification: • Robust Zero-Shot NER for Crises via Iterative Knowledge Distillation and Confidence-Gated Induction • Adaptive Log Anomaly Detection through Data–Centric Drift Characterization and Policy-Driven Lifelong Learning • ConFIT: A Robust Knowledge-Guided Contrastive Framework for Financial Extraction • Hierarchical Change Signature Analysis: A Framework for Online Discrimination of Incipient Faults and Benign Drifts in Industrial Time Series • Adaptive Evidential Meta-Learning with Hyper-Conditioned Priors for Calibrated ECG Personalisation • Hierarchical Adaptive Normalization: A Placement-Conditioned Cascade for Robust Wearable Activity Recognition 🌐 Want to build with EvoAgentX? Explore and contribute to EvoAgentX: 🔗 GitHub: github.com/EvoAgentX/EvoA… Ready for real-world AI workflows? Try the production-grade platform built on EvoAgentX: ⚡ x.com/HowoneAi— Autonomous agents, real work, zero config. We welcome builders, researchers, and developers to shape the future with us. 💥 #EvoAgentX #AgenticAI #AutoResearch #LLMAgents #MultiAgentSystems #AIWorkflows #HowOneAI #OpenSourceAI #AIforScience #ICAIS2025 #ReinforcementLearning #LaTeXAutomation #ResearchAutomation #AIAgent

Chop Chop

🫡 Invitation codes are limited

The more I think about it, the more this feels inevitable. Building a startup used to mean wrestling with tools, stacks, frameworks — an endless staircase of things you had to learn before you could even start. But AI is quietly dissolving that staircase. Not by making us “faster,” but by removing the parts that were never the real work to begin with. What’s left is the essence: a problem you care about, a vision for how things could be better, and the courage to say, “let’s try.” When the scaffolding falls away, creation becomes almost… pure. You speak, the system builds. You revise, it adapts. It feels less like traditional software development and more like shaping clay with your hands — immediate, intuitive, alive. So yeah, a 1-hour startup doesn’t sound crazy to me. If anything, it sounds like a return to what building was always supposed to be: not a technical marathon, but a conversation between your ideas and the world you want to create.

@mengzaiqiao @hayesdev_ Totally agree. The biggest shift isn’t “AI helps you build faster” it’s that AI becomes the builder. Humans just decide what to build.

@hayesdev_ I believe that in the very near future, you’ll be able to build a startup in just one hour — no need to spend an hour learning these tricks/AI stacks.

Ever wondered how a single person can ship a full agentic app? We’re letting a handful of early users into our beta. Jump in before the door closes. 👇

EMNLP Demo time 😎 We’ll be at Hall C3 from 14:30 to 16:00. Come see EvoAgentX — our framework for self-evolving LLM agent workflows. Yes, the “agents that actually get better over time.” If you’re at EMNLP and want to talk Agents, Reasoning, Agentic Memory, or Agentic Products with @mengzaiqiao & @JinyuanF pull up 👋 #EMNLP #LLMAgents #EvoAgentX

Love seeing EvoAgentX in action! 🔄 Real-world self-evolving AI, powered by creators.

The EvoAgentX Team is heading to Suzhou for #EMNLP2025! We’ll be presenting our demo paper — 👉 EvoAgentX: An Automated Framework for Evolving Agentic Workflows: arxiv.org/abs/2507.03616 EvoAgentX explores how AI agents can self-evolve, adapt, and collaborate to improve over time — moving beyond static workflows into continuous optimization loops. If you’re interested in agentic AI, self-evolving workflow systems, or workflow automation, come find us at the demo session tomorrow! 💡 Suzhou | @emnlpmeeting #AIagents #LLM #EvoAgentX #AIresearch #EMNLP #AgenticAI

Such amazing! HowOne AI !!

🚀 EvoAgentX × Telegram — Full Integration Now Live! We’ve fully integrated Telegram into the EvoAgentX workflow with the brand-new TelegramToolkit — bringing Telegram AI-native capabilities right into your agents 🔥 💡 What’s new? 8 powerful tools for your agents: ✅ Message Tools • fetch_latest_messages → Instantly fetch the latest messages from any contact, group, or channel • search_messages_by_keyword → Search historical chats by keywords • send_message_by_name → Send messages directly by name — no need to look up user IDs! • list_recent_conversations → Quickly list recent chats across users, groups, and channels ✅ File Tools • find_and_retrieve_file / download_file → Smartly locate and download Telegram files • read_file_content → Automatically read file contents (supports PDF and TXT) • PDF text extraction → Extract text directly from PDFs (powered by PyPDF2) ✅ Summary Tools • summarize_contact_messages → Generate conversation summaries with one call ✨ Why it matters • Simple name-based access to users, groups, and channels — no manual IDs needed • Built-in file reading and parsing, including PDFs • Seamless integration with LLM-driven agent workflows • Robust connection and error handling baked in With TelegramToolkit, your AI agents can now chat, search, summarize, and analyze — all directly inside Telegram. 📂 Example code: github.com/EvoAgentX/EvoA… 📖 Docs: github.com/EvoAgentX/EvoA…

🌿 EvoAgentX Demo — AI Feng Shui (the traditional Chinese art of spatial harmony) Advisor Workflow We tested a new multi-agent workflow built entirely in EvoAgentX — it takes a free-text home description and outputs a structured Feng Shui analysis & actionable layout plan 🏡 ------------------------------- 🧩 Workflow chain (1) input_analysis → interpret the user’s goal and living context (2) feng_shui_issue_identification → detect spatial & energy-flow conflicts (3) explanation_of_issues → reason about their impact on rest, focus, and balance (4) remedy_suggestions → propose renter-friendly, low-cost adjustments (5) summary_generation → assemble everything into a clear Markdown report ------------------------------- 📄 Sample output from the run • Home overview: two-bedroom apartment, 12th floor, southeast-facing, open-plan living + kitchen • Key findings: bedroom door alignment with entrance, cluttered balcony workspace, mixed energy zones, and lighting imbalance • Recommendations: ① Entry — add a light curtain or plant to buffer energy flow ② Living room — use area rugs to define zones and keep corners clear ③ Kitchen — declutter counters, improve light circulation ④ Bedroom — soften color palette, ensure bed is not directly in line with door ⑤ Balcony / workspace — organize surfaces, add greenery for vitality • Themes: energy control, calmness, focus, and overall spatial harmony • Budget: low to medium; all reversible for rental spaces ------------------------------- ✨ Why it matters This demo shows how EvoAgentX moves beyond question-answering — it decomposes a single natural-language goal into a multi-agent reasoning process, where each node interprets, diagnoses, explains, and refines before producing a human-readable report. The same structure can power workflows for travel planning, legal QA, marketing strategy, or design assistance — anywhere language can become process logic. ----------------------------------- 🌐 Open Source & Exploration If you're interested in multi-agent systems, automation workflows, retrieval-augmented generation (RAG), or AI-executable reasoning, you can experience it yourself — let AI help you build your own intelligent workflow. 📦 Project: 👉 github.com/EvoAgentX 📁 Example Library: github.com/EvoAgentX/Wond…

@ShipWithBhaskar Great contributions!

Just shipped Google Maps Toolkit into EvoAgentX (PR #141). Agents can now: Turn addresses ↔ coords Find “pizza near me” Compute travel time/distances Get timezones Location-aware AI just got easier. Repo: github.com/EvoAgentX/EvoA… #AI #Agents #OpenSource #Evoagentx #Google