a 551B model on a 64GB mac ~250GB of weights. 64GB of RAM. it runs anyway prefill 5.7 → 252 tok/s decode 3.3 → 16.3 tok/s 128K context the SSD is the new VRAM github.com/arnavprabhu/deeps…

Your AI agent is now just a phone call away. Mercury Agent, an open-source AI agent, can now call you directly on WhatsApp when something needs your attention. GitHub: github.com/cosmicstack-labs/…

Una simulación de Nueva York generada en tiempo real. Puedes andar, ir en bici, conducir coche o helicóptero. Con efectos de luz, tiempo real, tráfico, reflejos y más... Todo en menos de 800KB y de código abierto. → mrdoob.github.io/slop-newyor…

Get any website's full design system in one command, ready for Claude or Cursor to build from. It's a free open source tool called designlang (MIT license). How to use it: 1. Run: npx designlang stripe.com 2. It opens the page in a headless browser and reads the live DOM 3. You get 17+ files: design tokens, a Tailwind config, a shadcn theme, Figma variables and a paste-ready prompt pack for v0, Lovable, Cursor and Claude Artifacts 4. Hand those files to your coding agent before you ask it for a UI Extras worth knowing: - npx designlang site stripe.com crawls several pages and merges them into one system - npx designlang grade stripe.com gives the design a report card - It scores WCAG contrast and ships an MCP server for Claude Code, Cursor and Windsurf Needs Node 20+. Use it to learn from great design, not to copy someone's brand. github.com/Manavarya09/desig…

Claude Code, Anthropic's coding agent, shipped 2.1.309 on Oct 22. On macOS and Linux, sandboxed Bash commands ran the shell's startup files outside the sandbox before each command. That is fixed. If you turned the sandbox on to contain what the agent runs, update.

Memory engineering answers a different question from context engineering: what's true across runs, and who's allowed to write it. My new repo splits memory into scratch, episodic and durable tiers. Every note starts as a hypothesis, only a human or a verifier can promote one to durable, and retrieval runs on a budget. It also gives you a Memory Ready score from M0 to M3 that you can gate in CI. One Node command, no install, MIT. github.com/cobusgreyling/mem…

Interesting repo: shadcn-ui/lint (3250 stars). It is an agent-first linter for Tailwind design systems. It is written in TypeScript. You write the rules. The linter checks them. The goal is simple. Your design system rules become machine-checkable. When an agent breaks a rule, the error says what is wrong. It also suggests a fix. The fix comes from your components, your variants and your theme. It works with your existing design system. No rewrite. It needs Tailwind v4. shadcn/ui is not required. It runs on ESLint and Oxlint. It works with React, Svelte and Vue. Here is why it matters. A TypeScript type can block a prop. It cannot explain the design intent. The readme uses a Button that permits margin and width, but controls its own padding. TypeScript gives you TS2353: padding does not exist in type Pick. The agent learns that padding is blocked. It does not learn how to size the Button. The linter gives the second part. The error says the Button owns its spacing. It says to use a size (sm, lg), or margin here, or gap on the parent. It says to add a size in components/ui/button.tsx only if the design calls for one. That is a fix an agent can apply. What you can do: - Write contracts per component. One example lets CardTitle change layout and typography, but not font-*. Another lets CardContent change layout and spacing. - Keep values on your theme scale. Pair shadcn/no-restyle with shadcn/no-arbitrary-values set to "error". - Set it up with one prompt to your coding agent: Read github.com/shadcn-ui/lint/bl… and set up @shadcn/lint in this project. The config is small. One example sets shadcn/no-restyle to "error" with layout permitted. It adds a contract for ^Button$ that lists w-full, mt-* and mb-*. That is the whole rule. This is a good pattern for a team that runs agents on a shared component library. The rules live in config. They do not live in the agent prompt. The agent gets the same answer every time.

🚨 ClaudeDevs dona Agent Skills a la Agentic AI Foundation ClaudeDevs ha anunciado la donación de Agent Skills a la Agentic AI Foundation, la misma fundación a la que donó MCP. Desde diciembre de 2025, más de 70 plataformas de agentes han integrado el estándar, entre ellas ChatGPT, Cursor, GitHub Copilot y VS Code.

I trust an agent's explanation of a codebase more when it has to draw it. I recently shared native-surface, a library that renders react native apps in the browser without the DOM. It's tricky to explain in text, but a diagram makes it clear. Stop asking your agent to only explain the code. Ask it to draw a walkthrough.

running Recursive Flint 40B-A2.5B on a single RTX 3060. 12GB VRAM, 16GB RAM. (config below) 256K context. ~46 tok/s decode, ~675 tok/s prefill. brand new coding moe, built from scratch in under 4 months. llama.cpp support isn’t merged yet, so this runs on PR #30928 cherry-picked onto master. only 14 of 49 layers are full attention, so 256K ctx costs just 3.8GB of q8 kv. benchmarks next! config: llama-server -m flint-preview-UD-Q3_K_XL.gguf -ngl 99 --n-cpu-moe 33 -c 262144 -fa on --jinja -np 1 --cache-ram 0 --cache-type-k q8_0 --cache-type-v q8_0 -b 2048 -ub 1024 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --repeat-penalty 1.0

Your coding agent shouldn't guess CVE facts. Crowbar puts NVD, CISA KEV, EPSS, Exploit-DB, Nuclei, and Metasploit in one local SQLite file and serves it to Claude Code, Cursor, and Codex over MCP. Offline, open source, one command: npx crowbar-cli setup github.com/kushvinth/crowbar

Anthropic’s coding agents pushed daily merges from about 2,500 to more than 30,000. Its engineering report shows what happened to the machinery around the code: checkouts, caches, history queries, and testing.

🚨 ¿QUÉ ESTÁ PASANDO CON OMNI 1.1? en X en inglés se repite el mismo post una y otra vez: "este video me costó $0". lo que no te dicen: 1. gratis es el plan de prueba. en API cada clip de 6-10 s nos sale ~$0,4-0,6. 2. el video es tan bueno como la foto de entrada. foto de plástico = video de plástico. 3. lo que delata es la voz: si no pides micrófono de celular y pausas imperfectas, suena a locutor. 4. hoy hicimos 2 clips en español chileno (feria y asado) al primer intento. barato sí. magia no. guárdalo 🔖

how do i check what 10-15 agents did before it reaches a client? glad you asked spec driven development (SDD) + orchestrator-workers, with my own manager layer on top 1. spec first client gives me the goal. i turn it into one roadmap, split into milestones. each milestone breaks into features, each feature gets its own pedantic PRD. one PRD, one agent. this part alone can take a day or two and it's worth every hour 2. the hierarchy orchestrator -> manager (milestone A) --> workers A1, A2, A3 -> manager (milestone B) --> workers B1, B2, B3 and so on orchestrator checks if anything changed since i wrote the roadmap (clients love changing design on the fly), then spins up one manager per milestone. workers can be any grunt model, sonnet, deepseek, qwen, some flash model, whatever fits 3. why the split these runs go 3-10 hours. one session holding everything hits its limit, compaction eats specifics, and suddenly the thing you talk to forgot the details so managers only report up in one-liners. keywords, max signal, zero noise. the orchestrator stays lean for hours, and even at 600-900k tokens its context is tool calls and keywords, not rot. that's the session i ask for status or throw quick changes at 4. review at every layer worker done -> manager checks it against the PRD, wrong = back it goes (evaluator-optimizer) milestone done -> separate reasoning model, fresh context, reviews only the code (writer/reviewer) all done -> one model checks the whole thing front to back + e2e, critique goes straight to whoever wrote it 5. receipts terminal + local git worktrees. every change lands in history, so when something breaks i point at the exact agent and the exact work segment I dont trust agents, so they get made verifiable and then critically checked

Monty can pause agent-written code, turn it into bytes, and resume it later ⏸️ - dump() serializes the whole interpreter - load_snapshot() resumes it in another process or machine - Wait hours for a human approval with nothing running pydantic.dev/docs/monty/conc…

Google got AI agents up to 11.8 points better without touching the model. Harness² freezes the weights. It only edits the prompts, tools and memory around the model, tests each edit on the same task, and keeps what worked. Open source. What's one prompt or tool change that fixed your agent?

Node, Python, variables de entorno y tareas en un solo archivo: mise.toml. Eso es mise (MIT, 35k ⭐, en Rust): - 1.000+ herramientas, una versión por proyecto - carga tu .env por proyecto - mise run para build y test ¿Sigues con un gestor de versiones distinto por lenguaje?

🕷️ Scrapling tip: pierce_shadow=True in v0.5 brings open Shadow DOM content into the response. Read it with normal CSS or XPath, including text assigned to slots. No custom JavaScript to walk shadow roots. Closed roots stay out 🎯