Introducing Eigent — the first multi-agent workforce on your desktop. Eigent is a team of AI agents collaborating to complete complex tasks in parallel. It is your long-term working partner with fullly customizable workers and MCPs. Public beta available to download for MacOS, Windows. 100% open-source on Github. Comment for 500 extra credits.

Very late to the party. When Thinking Machines Lab released their blog on On Policy Distillation, my first reaction was that it should be just like DAGGER from 15 years ago: arxiv.org/abs/1011.0686. I finally had time to read the blog today and sure enough, they mentioned DAGGER. Actually, if you have access to an expert, there are quite a few on-policy imitation learning approaches you can try. Here’s one incremental idea that really works from an obscure paper we published at 2018. It is basically an on policy multi-teacher imitation learning method that chooses the best teacher for each state based on their value functions. We found that it’s possible to imitate multiple imperfect teachers simultaneously, and eventually outperform all of them. Naturally, the more diverse those teachers are, the better. Would be interesting to reproduce this in the LLM era: arxiv.org/abs/1803.01129v3

Automatic token management in CAMEL

brainwash your agents. context engineering doesn't have to be hard, there are so many low-hanging fruits. just keep the memory a holy place and drop the bs messages I just wrote a blog post on how we do it at @CamelAIOrg. these are simple to implement, must-have techniques for apps that use agents, which can optimize the accuracy and the cost without crazy code changes. 🎁 BONUS: I created a number of bite-sized issues that you can get on right now and start your open-source arc. just open a PR, or help review one. read here: shorturl.at/zyCv7

we shared three context engineering techniques implemented in the @CamelAIOrg framework that keep agent memory clean and context sharp: - context summarization - workflow memory - tool output caching blog: camel-ai.org/blogs/brainwash…

This project perfectly matches how I've long envisioned AGI: agents that self-explore, self-annotate, and self-verify all the training data they need! There are absolutely no human annotations, no demonstrations, and no pre-defined tasks, and the method can work on any app or platform to improve your agents. When data synthesis becomes a trend in training LLMs, what is the status quo for agent data? The greatest challenge is still the lack of grounded action sequences and environmental state info. Common practices rely on human supervision or synthetic data generation by executing pre-defined tasks, which are either resource-intensive or fail to ensure data quality. Moreover, these methods suffer from limited data diversity and significant gaps between synthetic data and real-world environments. Ideally, we want a manual-free data pipeline for synthesizing GUI agent trajectories, free of human annotations, demonstrations, and pre-defined tasks. However, synthesizing valuable tasks for GUI/OS environments that cover a broad range of GUI functionalities is inherently difficult, because neither humans nor LLMs have full prior knowledge of the environment. We then asked ourselves: given an app or a website, how do humans explore it? Beyond common knowledge, humans explore by interacting—trial and error! This reflects the intuition that actual usage uncovers hidden functionalities and states beyond preconceived tasks. We introduce Interaction-Driven Functional Discovery, a rule-based exploration process in OS-Genesis that systematically traverses dynamic GUI environments by interacting with UI elements (e.g., clicking, typing, scrolling). This approach uncovers diverse functionalities without relying on predefined tasks, collecting interaction triples—actions, pre-action states, and post-action states—while exploring the breadth and depth of a platform. We then use Reverse Task Synthesis to transform these observed interaction triples into meaningful low-level and high-level task instructions. It first generates low-level tasks (e.g., “Open the Calendar app”) and further derives corresponding high-level tasks (e.g., “Add a reminder to post about our paper on X today”). A trajectory reward model ensures the quality of the generated trajectories. This process of “interactions to tasks” is essentially the opposite of “pre-defined tasks to trajectories,” offering greater flexibility.

Introducing Firecrawl v2.5 - the world's best Web Data API 🏆 We now have the highest quality and most comprehensive Web Data API powered by our new Semantic Index and custom browser stack. See the benchmarks and technical deep dive below 👇

Memory is what makes us human. It's also what makes AI truly intelligent. @mem0ai has raised $24M to build the universal memory layer for AI. Thousands of teams in production. 14M downloads. 41K GitHub stars. Intelligence needs memory & we're building it for everyone. More👇

We introduced workflow memory into @CamelAIOrg. Thanks @Hesamation for leading this effort!

Thrilled to release new paper: “Scaling Latent Reasoning via Looped Language Models.” TLDR: We scale up loop language models to 2.6 billion parameters, and pretrained on > 7 trillion tokens. The resulting model is on par with SOTA language models of 2 to 3x size.

🐫 We’ve partnered with @qdrant_engine on their Essentials Course, all about production-grade vector search. Check out our module to learn about: - Multi-agent system architectures - Agentic RAG patterns and best practices - Agent collaboration and communication - Building autonomous AI systems with Qdrant - Auto-Retrieval with CAMEL for automated RAG processes View the full course: qdrant.tech/course/essential… View our module: qdrant.tech/course/essential…

Thanks for building on top of CAMEL!

😫Struggle with preparing presentation video before Deadline (such as NeurIPS)? 🔥🔥Thrilled to share our latest work — Paper2Video — automatically generates presentation videos from papers!! 🚀🚀Just provide your paper➕a portrait photo➕a short audio sample — our Paper2Video will create a full presentation video for you. Try Paper2Video and let us know your thoughts!! 💻 GitHub: github.com/showlab/Paper2Vid… 🌐 Website: showlab.github.io/Paper2Vide… 📜 arXiv: arxiv.org/abs/2510.05096 🤗 HF Dataset: huggingface.co/datasets/Zayn… 🤗Daily Paper: huggingface.co/papers/2510.0… 🎉Paper2Video is accpeted by SEA Workshop @ NeurIPS 2025 sea-workshop.github.io and will be presented this year Decemebr! 🙏Kudos to the amazing @zayn42682 and @MikeShou1 Our work is built on top of the multi-agent framework @CamelAIOrg by @guohao_li. Huge thanks @_akhaliq for sharing our work!

Introducing CAMEL browser toolkit. Our built-in browser toolkit in CAMEL framework. Most open source agent browser tools suck. So we built our own and use it in @Eigent_AI product. In this blog post, we discuss how we build our browser toolkit - the action and observation space for agent browsing. - Multi-mode operation - Typescript framework integration - Enhanced element identification - _snapshotForAI and ARIA mapping mechanism - Enhanced stealth mechanism - Tool registration and screenshot handling Blog link👇

Most agent frameworks look the same. What makes CAMEL-AI different? Our native toolkits. In this blog we show how the Hybrid Browser Toolkit lets agents actually see and interact with the web and plug straight into MCP. Detailed Blog👇

SAP Purchase Orders without the clicks. Our latest blog shows how Eigent’s multi-agent AI workforce can log in, create, and submit a PO in S/4HANA, while you focus on more important work. End-to-end ERP automation, right from your desktop. Full guide 👇

🚀 Introducing DeepSeek-V3.2-Exp — our latest experimental model! ✨ Built on V3.1-Terminus, it debuts DeepSeek Sparse Attention(DSA) for faster, more efficient training & inference on long context. 👉 Now live on App, Web, and API. 💰 API prices cut by 50%+! 1/n

Wait, did not see this in my timeline at all. - 1T open source reasoning model with 50B activation - Solid results on math and coding Are people not excited about new open source model releases any more?

Ever wondered if an AI can actually scale creative tasks with just one prompt? We asked Eigent to generate 200 HTML games on different topics, split them into 200 subtasks, run everything in parallel, and package the output into a single zip. Eigent handled the whole workflow: > Decomposed the task automatically > Ran 200 subtasks in parallel > Delivered one zip with 200 playable games Try this out today. Eigent shows how a local AI workforce can take one idea and scale it hundreds of times over.