AI agents are becoming commodities. Research taste isn't. The real moat won't be another model. It will be the research harness you've spent months or years evolving. One that accumulates your judgment instead of repeatedly asking for it. One that gets better after every paper, every experiment, and every failure. That's what we're building at @EvoScientist. Today we crossed 4.5k GitHub stars. 🚀🚀🚀 Thanks to everyone who's been part of the journey. More to come.
From AI power to the power to decide what's legal. Welcome to the next stage of the AI race. x.com/demishassabis/…
Dude, this kills the game. It knows almost everything I care about — my research rhythm, repo changes, papers, meetings — and turns them into proactive scheduled briefings. Actually sensing my workflow. 👏👏👏 x.com/openai/status/…
Research. Remember. Evolve. Watch your research graph evolve in EvoScientist. 🧠🌱
Research. Remember. Evolve. Watching a research graph come alive. 🧠🌱
A few months ago, I wrote: "Harness Vibe Research: Autonomy Is Earned, Not Installed." It wasn't just a slogan — it was a goal. Today, I finally feel like I have a small but real answer. While I was away attending a conference, I left EvoScientist running in its own environment for just 7 days. It started with some of my research background and preferences, but every day it continued reading the latest papers, technical blogs, and reflecting on its own execution. What surprised me most wasn't the number of papers it read. It was watching its local knowledge base evolve from isolated observations into an interconnected research wiki. Nodes gradually became clusters around research directions I genuinely care about, while entirely new exploration paths emerged on their own. Even more interestingly, the system didn't only accumulate research knowledge — it also learned from itself. Mistakes made by the harness during execution were distilled into procedural memory, preventing the same failures from happening again. Tool usage became more reliable, paper-search strategies improved, and reusable research skills continued to evolve with every run. At that moment, I realized the system wasn't only learning what I care about. It was continuously improving how it learns, and how it carries out research by itself. Knowledge doesn't disappear after a session. It compounds across runs and is passed forward into future exploration. To me, this is what a self-evolving research system should look like — not just an AI assistant with memory, but a research harness that becomes a better researcher over time. Looking back, it feels like that article from a few months ago wasn't just an idea anymore. It finally started to happen. I'm excited to share much more when I'm back. This has been one of the most fascinating experiments I’ve run so far. 🚀
Strongly resonate with this. At EvoScientist, we believe the next stage of AI agents is not only about stronger base models, but also about the harness around them: memory, tools, workflows, skills, artifacts, and self-improving loops. This is exactly what we are building for AI auto-research. Our memory is continuously evolved through an asynchronous harness layer across every turn and every agent session. Each interaction can update long-term memory, refine the project wiki, and accumulate reusable knowledge for future research. We are also exploring self-evolving research skills, where agents improve how they plan, search, code, debug, write, and evaluate over time. Auto-research is a great testbed because it combines long-horizon exploration, uncertain objectives, noisy results, and multi-agent coordination over persistent artifacts. EvoScientist is our open-source attempt to build a harness for self-evolving AI scientists. #EvoScientist #AIforScience #AgenticAI #HarnessEngineering #OpenSource
Claude Science validates a direction we deeply believe in: research AI is moving from chat assistants to full research ecosystems. x.com/evoscientist/s…
🎉 Heading to #ACL2026 in San Diego 🇺🇸! Excited to share that I’ll be presenting our latest work at ACL 2026: 🩺 CCD: Mitigating Hallucinations in Radiology MLLMs via Clinical Contrastive Decoding w/ @mengzaiqiao, @jakelever0, and @EdmondSLHo. 📍 Poster Session B 🗓️ 14:00–15:30, Sun, July 5 📄 Paper: t.co/XpNb0RCjFP 🌐 Website: t.co/DhmX0zJkpy 💻 Demo: t.co/bpjRMTnLDw 🚀 CCD is a training-free and retrieval-free decoding framework for radiology multimodal LLMs. By integrating expert clinical signals directly during decoding, CCD helps reduce hallucinations and improve clinical consistency in radiology report generation. If you're attending #ACL2026, feel free to stop by and chat — I'd love to connect with researchers, builders, and collaborators working on MLLMs, clinical NLP, radiology AI, and autonomous research. #ACL2026 #ClinicalNLP #RadiologyAI #MLLMs #MedicalAI #AIforScience
This makes a lot of sense. We’re moving from standalone AI systems to self-evolving agent ecosystems, where context is no longer ephemeral but becomes part of a compounding intelligence layer. Auto-research is just one good example of that broader shift. x.com/engramlab/stat…
@codewithhajra Thanks! That’s exactly the direction we’re trying to push towards.
@aaliya_va exactly!
@Xudong07452910 cheers!
🚀 Shipping the new EvoScientist WebUI. Designed to make Vibe Research feel more natural: 🧠 EvoMemory 🔬 Research Skills 🤖 Multi-Agent Workflows 📂 Workspace Management Helping researchers spend less time managing information and more time exploring ideas. #OpenSource #AI #AgenticAI #Research #VibeResearch #AIforScience
📄 Paper: arxiv.org/abs/2509.23379
📦 GitHub: github.com/X-iZhang/CCD
