We open sourced BrowserSkill, a bridge between your agent and your actual browser. most tools give the agent a blank browser. We let it borrow a tab from yours, then hand it back. > login state is already there, it just works where you're signed in > captchas and confirmation dialogs come back to you, then it continues > it's a CLI, not an MCP server => any agent that can run a shell can use it, and you see every call it makes one thing that's easy to miss: the agent asks before borrowing a tab, and that switch lives in your browser settings, not in a flag, so it can't be talked around. one line to install, works with Cursor, Claude Code, Codex, Hermes, Openclaw, CodeBuddy, WorkBuddy. Everything runs locally, MIT.

Can we get a redo button while we're at it? Secure VM put agent isolation on everyone's mind. Rollback is the other half of safety. When a step goes wrong, you want to undo it. @CubeSandbox is open source. Agents go off-script all the time, so each task gets its own kernel and untrusted code never touches your stuff. Your credentials sit in a vault and get injected at the gateway, never inside the sandbox. And when a step still goes wrong, you roll back to a snapshot instead of starting over. 🤯repo:

We open-sourced CubeSandbox for a reason: so builders tell us what's broken. They did: → Cross-node pause/resume. Pause on one node, resume on another. Snapshot restore and stability both improved. → Component failure used to take the cluster down. CubeMaster multi-replica, template service split out. No single point of failure left on the core path. → LLM timeout: 60s proxy cutoff was killing slow-first-token models. Now 2h. Next thing builders told us @CubeSandbox : pause/resume memory cost is too high, on it.

The open-source WeKnora (MIT, 22k stars) just shipped v0.8.0. A new step from RAG toward agentic RAG: a self-remembering, safely sandboxed knowledge agent platform. What's new: 🧠 Cross-session memory: profile, preference, fact, task and interest, each proposed for your confirmation before it sticks. 🔐 Sandboxed by default: Skills used to run in the host process; that backend is gone. Each session now runs in its own sandbox on Docker, E2B, or Cube, with configurable network egress policies. 📦 Skills as assets: a tenant-wide catalog, not scattered scripts. Pull them from ClawHub, SkillHub or any git repo, versioned and verified to load before anything runs.

Quick one: DeepSeek V4.1 Flash lands on WorkBuddy/CodeBuddy, co-launched exclusively in China, discounted for two weeks. TokenHub, ima,​ and Marvis are live day 0 as well. Well played, @deepseek_ai 🫡 V4.1 Flash is a really solid model, major gains across text and agent performance, with native multimodal visual understanding. Yes, we have our own models too, but open beats owned, it always has.

Fun fact: your agent is quietly wrong about the past. Outdated facts aren't deleted, just shadowed to the back. RoMem fixes that. Drop-in reranker, three lines to plug in. The key ideas: > Spin, don't delete Outdated facts rotate out of alignment at query time. Nothing erased. > Some relations expire faster A pretrained gate scores how fast each relation changes. born_in → near zero. works_at → fast. Zero-shot to unseen relations. > Time as rotation, not a label One continuous function. Any timestamp. Including dates never seen in training. 2–3x better at temporal reasoning, with zero drop in static memory. Joint research from Edinburgh, LIGHTSPEED and St Andrews. EMNLP 2026 Main. Paper: Code:

A cache is only worth what it hits. The community's having a KV cache moment. Here's the corner we work in: In long-context serving, a cache hit can still leave the GPU waiting for data. When KV lives outside GPU memory, how you bring it back matters. That's the problem we set out to solve. FlexKV restores it layer by layer: earlier layers compute while later layers load. Prefetching starts transfers early, while asynchronous writeback helps overlap cache I/O with inference. Making those hits faster goes hand in hand with making more of them possible. FlexKV compresses KV losslessly, expands cache capacity with CPU RAM, SSDs, and remote storage, reuses prefixes across the cluster, and routes requests to wherever the cache already lives. It sits under your inference engine, so there’s nothing to rewire. Works across SGLang, vLLM, TensorRT-LLM, and Dynamo. Up to 70% lower TTFT, +16% QPM.

We used TeamAI-CLI internally at Tencent since March. Open sourced it. It turns team knowledge into one git repo, so every agent works from the same handbook. what it does: > git-based: skills, rules and docs live in one repo, changes go through a merge request > hook-triggered: merged changes land on everyone's next session > each learning earns confidence from real usage, strong ones surface first, weak ones sink > works with claude code, codex, cursor, opencode, codebuddy and workbuddy One person's hard-won workaround can now become the whole team's default.

1B requests a day in Weixin. Now open source. WeMM-Embedding is built for real traffic and verified in real use. It reads text, image and video in the order they show up. The 9B ranks #1# on MMEB-v2 (80.6); the 2B keeps 98.7% at just 256 dims. Truncate to 64 dims for fast recall or 2048 for fine ranking, no retraining. Try it:

Using Taxis to explain the Sherry method: Imagine a taxi with exactly four passenger seats. In the old way of doing things, sometimes a seat has a person, and sometimes it is empty. But even when a seat is empty, the taxi still takes up the same amount of space on the road and burns the same amount of fuel. The empty seat is WASTED SPACE. The Sherry method introduces a clever new rule: Every taxi must carry exactly THREE passengers, leaving exactly one seat EMPTY. Because you know there will always be exactly one empty seat, you can use the position of that empty seat to send a secret message to the taxi driver: - If the front seat is empty, it means "We are going North." - If the back-left seat is empty, it means "We are going South." - If the back-right seat is empty, it means "We are going East." Etc. Now, the empty space isn't wasted anymore. That 0 value is doing a job. @TencentHunyuan

The agent race is moving from models to context. Context isn’t just what you type. A connected agent gets the whole picture, not just the words. WorkBuddy just made it real, 60+ brands across hardware and software, including @insta360 , @RokidGlobal , and @meituofficial . Apps run on a base. Agents just got an open one. 👀

1.5TB → 214GB. Seven times smaller, barely a dent. That's Hy4 preview. The trick is Sherry — our quantization method that packs weights down to 1.25 bits each. (The image shows how.) It also unlocks a new way to run it: stitch the GPUs you already have across machines, and they work as one. GGUF (quantized): Original model:

Introducing Hy4 preview ◆ Open weights: 770B MoE, 49B active, 1M context ◆ Built for real work: code, docs and analysis, scientific research ◆ Hy4 preview helped optimize its own training pipeline, kernels included ◆ We ship upgrades, not price hikes Free in @WorkBuddy_AI for the next two weeks Blog： GitHub： Get WorkBuddy：

While everyone's debating when AI leaves the screen, a hundred or so driverless trucks have already logged 45 million kilometers on highways. That's KargoBot, running L4 across 20+ trunk routes, 1.4 billion ton-kilometers of freight hauled so far. The data never stops coming, and at PB scale raw compute alone doesn't get a model to its next version. Our bit: — compute clusters for training and inference, built to scale with the fleet — object storage with a caching layer in front, so training isn't waiting on IO — TI-ONE, our ML platform, holding the iteration pipeline together Software agents retry for free, a 40-ton one doesn't. That's what makes infra fun to work on. What's the part that breaks first when your agent leaves the screen?