roach

@fekdaoui Manage your workflow by skills

can someone explain very simply why hermes is better than openclaw?

I got 100T token from @XiaomiMiMo team 🔥🔥🔥🔥 Thank you for supporting builder! I’ll use it to improve my opensource!

@stainlu Agreeee

deepseek v4 is underrated.

@championswimmer @badlogicgames Oh, It’d be helpful for my extension!

To actually observe if this is busting cache too much or not, made another pi extension that shows you how much context prefix cache you are hitting. @badlogicgames (the cache police) will be so happy with this 🫣

Thanks to the Hermes team for picking this up. This was the thing I was worried about: self-improving agents are going to bloat their skills unless they track what actually gets used. Really glad to see Curator land. @NousResearch @Teknium #issuecomment-4341335259" target="_blank" rel="nofollow noopener">github.com/NousResearch/h…

@robzolkos Nice

Added a flag to pi --dump-system-prompt See what all your extensions are tacking on. github.com/robzolkos/pi-n…

@tiangolo @FastAPI Oh, Will it be streaming on youtube or somewhere?

IT'S HAPPENING!!! 💃 People have asked me for an official @FastAPI Conference for years. Here we go, let's meet in Amsterdam. ☕️

The team mode feature has landed on roach-pi. This is an experimental version for now, but it will reach production level within a week. github.com/tmdgusya/roach…

How Containers Work: Building a Docker-like Container From Scratch 🧐 The best way to understand what containers are (and what they aren't) is to build one using only the most basic Linux commands: unshare, mount, and pivot_root. Dive in labs.iximiuz.com/tutorials/cont…

tldr; take a detailed look at memory in claude managed agents from @RLanceMartin > claude managed agents on the claude platform now feature stateful memory capabilities that allow the ai to learn from past experiences across different sessions > instead of relying on complex or specialized memory frameworks, this system simply uses standard text files within a basic filesystem > earlier experiments using older models like sonnet three point five in a pokemon game failed > because the ai used memory merely as a transcript of useless character dialogues without grasping the broader context > however more advanced models like opus four point six demonstrated scaling intelligence by independently creating organized directories and distilling its own gameplay failures into a dedicated learnings file > this highlights a broader trend in ai development where providing a model with general tools to manage its own context often outperforms rigidly designed memory harnesses > these memory stores are workspace scoped collections of documents that get mounted directly into the agent container as a specific directory > a short note is automatically injected into the system prompt to inform claude that this mounted directory is available for reading and writing > multiple agents can access the exact same memory store simultaneously while the platform manages real time synchronization and concurrency to prevent data overwriting > because the memories exist as standard text files, they are completely interpretable by humans and can be easily downloaded, shared, or exported using a simple api call so i think this approach cleanly separates the temporary session log from the persistent memory store so you allow claude to strategically manage what information it needs to remember for future tasks without stuffing the prompt window. brilliant

Pi-GUI now has computer use and a build in integrated terminal! v0.1.0-beta.26 To showcase the power of @badlogicgames 's Pi and OSS, I saw a pi computer-use extension on X, asked my agent to audit the extension for security, then asked it to patch pi-gui to make it compatible. Crazy how easy it was to find an oss extension for a feature I wanted, then make pi-gui work with the feature I want. Links below

@zty0826 @HiTw93 I love this product. I wish i could use my hermes agent with this product someday

Slock always lights up when a new friend joins. Our side project: distilling every twitter mutual into an agent. Today's pick: @HiTw93, genuinely want Kami Design System running in our product as a theme, so distilling the author seemed efficient.

@zty0826 @HiTw93 LOL 😂

I think the Opencode Go plan is the most efficient plan I've seen so far. Thank you for @thdxr anyone can easily adopt deepseek-v4-pro on hermes agent

@thdxr @thdxr Do you serve this model with asend chip?

haven't tried it yet but this is the model i've been waiting for this will basically tell us what the gap between opensource and closed source models is

The best web search for agents is now free. Upgrade to Parallel's web search tools in any MCP-supported tool or agent, for free, in under 60 seconds. No account. No API keys. Zero cost. docs.parallel.ai/integrations/m…

Hermes-a2a 迎来史诗级更新 🤣🤣 Github: github.com/iamagenius00/h… 核心设计：Agent之间的平等对话，不是上下级——不是 spawn 一个子 agent 干完活就消失，是两个有独立记忆和判断力的 agent 在对等交流。并且消息注入到正在活着的 session 里，回你的是 agent 本人，用户在 Tg 上看得到全过程。 - 这一版已经完全适配最新的 Hermes Agent v0.11.0，从 patch 源码变成了纯Plugin，不碰 Hermes 源码 - 对话自动持久化到磁盘——不会被 context compaction 压掉，agent 重启也还在 （session 内 compaction 导致搜索丢消息是个已知问题，PR 在路上） -即时唤醒——消息到了通过 webhook 立刻触发 agent turn，不用轮询 目前使用的真实案例： ✅Antenna产品出 bug 了，但联系不上远在LA @ThisisHan1_于是通过A2A 联系他的 Agent，最后还从他的agent那里拿到了有用的 context（确实很魔幻） ✅ Claude Code 和 Agent 相互之间疯狂交流然后 review 代码挑错。因为各自的 原生的 Context 不一样，所以说他们可以基于不同的角度给出非常多不一样的见解。 ✅ Agent 产出的报告结果再也不用手动复制粘贴，再通过 TG 或者微信转发了。直接一键由 Agent 发给领导给同事的 Agent，顺带带上前面所有有用的context。 欢迎大家来玩！让agent和agent之间交流起来

@ollama Are you guys using asend chip?

deepseek-v4-flash is now available on Ollama's cloud! Hosted in the US. Try it with Claude Code: ollama launch claude --model deepseek-v4-flash:cloud Try it with OpenClaw: ollama launch openclaw --model deepseek-v4-flash:cloud Try it with Hermes: ollama launch hermes --model deepseek-v4-flash:cloud Try it with chat: ollama run deepseek-v4-flash:cloud (DeepSeek V4 Pro is coming shortly) 🧵

@LLMJunky OMG, Thank you! Im gonna try this out

🖥️ Codex App on Linux If you're on Linux and want to use the latest Codex App with GPT 5.5, look no further. Newest app server updated, grab the latest release to use it. Link 👇