GitHub热门项目 | | Stars: 1,995 | 6 stars today | 语言: JavaScript

GitHub热门项目 | Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills. | Stars: 27,552 | 180 stars today | 语言: Python

GitHub热门项目 | "OpenSpace: The Skill Management Layer for AI Agents" -- https://open-space.cloud/ | Stars: 7,129 | 91 stars today | 语言: Python

GitHub热门项目 | A curated list of awesome libraries, packages, strategies, books, blogs, tutorials for systematic trading. | Stars: 9,199 | 113 stars today | 语言: Python

600 MiB of headroom My RTX 4070 was running Qwen 35B beautifully after the --cpu-moe trick from a previous run. The tokens/sec were where I wanted them. VRAM sat at 11,714 MiB out of 12,281 — 95% full. That leaves 600 MiB. Not enough for a serious agent. The context window I was giving llama.cpp was -c 4096 . Fine for chat. Not fine when a Claude Code-style agent hands the model 12,000 tokens of tool definitions before it says hello. I wanted -c 32768 . That's an 8× jump. And the memory that grows with context length is the KV cache. Multiply the cache by 8 with 600 MiB free, and llama.cpp dies during warm-up. I know because I tried it first. What actually sits on the GPU After offloading the MoE experts to CPU (the previous chapter's trick), the GPU is holding two things: The attention weights and non-MoE parameters The KV cache — a running record of every token the model has already read The first is fixed. The second grows linearly with context length. Double the context, double the cache. -c 4096 → -c 32768 doesn't just want 8× more tokens processed, it wants 8× more cache resident in VRAM the whole time. There is no room. So the cache itself has to shrink. Two flags llama.cpp takes two flags for KV cache dtype: llama-server -m qwen35.gguf -ngl 99 --cpu-moe -c 32768 \ -ctk q8_0 -ctv q8_0 -ctk is the Key cache, -ctv is the Value cache. Default is f16 (16-bit). q8_0 cuts each in half. Halving both means the KV cache footprint drops by roughly 50%. That freed-up VRAM is exactly what I need to make the context 8× bigger without touching the model weights. The measurement Same prompt, same seed, two runs — one at f16 KV, one at q8_0 KV: KV dtype Max -c I could allocate Tokens/sec (decode) Perplexity delta f16 (default) 4096 ~34.6 baseline q8_0 32768 ~34.1 negligible in my tests The speed loss is inside noise. The context is 8× longer. The quality drop I could not tell apart from run-to-run variance. Community measurements agree: symmetric q8_0 KV lands somewhere unde

Granola debutes an Apple Watch app for in-person notetaking

We run a React app and an Angular admin panel at work. Same company, same brand, and on paper the same design. On screen it was a different story. The React button had a 6px radius; the Angular one had 4px. The focus rings were two slightly different blues. Nobody noticed until somebody did. And every time design changed a token, someone got to hand-port it into two codebases. Twice the work, and it still drifted. So I went looking for something that treated both frameworks as equals. The React kits don't speak Angular. The Angular ones don't share a look with anything on the React side. Nothing let me define the design once and have it show up, the same, in both. So I built bpdm/ui . One rule: the look lives in tokens I gave myself one hard rule: nothing about how a component looks is allowed to live inside the React or Angular code. Colour, spacing, radius, the easing on transitions, all of it sits in @bpdm/tokens as plain CSS variables, and both framework packages just read from there. The component owns structure, behaviour, and the accessibility plumbing. The look comes from the tokens. @import "tailwindcss" ; @import "@bpdm/tokens/tokens.css" ; Change one token and both frameworks move together. There's no "now go sync the Angular theme" step, because there's only one theme to sync. Four ship in the box (two light, two dark). Override the variables and you've re-skinned all of it. The same component, twice React: import { Button , Badge } from " @bpdm/ui " ; export function Example () { return ( < Button variant = "primary" > Get started < Badge appearance = "soft" > New </ Badge > </ Button > ); } Angular: import { Component } from " @angular/core " ; import { BpdmButton } from " @bpdm/ng " ; @ Component ({ selector : " app-root " , imports : [ BpdmButton ], template : `<button bpdmButton>Get started</button>` , }) export class App {} Same padding, same radius, same focus ring. The accessibility isn't literally shared code: Radix does that work on the React s

This year I've built more MCP servers than I meant to, across three pretty different shapes: vellum : a self-hosted server over a folder of markdown, so my agent has a memory that's just files I own. (Open source, MIT.) Metrifyr : a marketing-data server that puts GA4, Search Console, AdSense and Tag Manager behind one connection. A read-only gateway over a company's internal user-data stack, federating half a dozen internal services behind one audited MCP surface. Personal, product, enterprise. Totally different data, totally different stakes. And yet the things that made each one good , versus a flashy demo that falls over the moment a real agent uses it in anger, were the same every time. Here are the seven that keep recurring. 1. Every tool you expose is a tax on the agent's context This is the one nobody warns you about. Every tool definition (name, description, JSON schema) gets loaded into the model's context on every single request , whether or not it's ever called. Twenty tools of boilerplate and the agent is reading pages of plumbing before it touches your data. So the design pressure runs the opposite way from a normal API: fewer, broader tools beat many narrow ones. vellum has 15 core tools and I fought to keep it there. If a tool doesn't earn its slot in the context window on most requests, it shouldn't be a tool. 2. Tools are for acting. Resources are for reading. The corollary to #1. Most servers expose everything (including "read this thing") as a tool. But MCP has a better primitive for reading: resources . In vellum, every note is a resource at a stable URI ( vellum://note/projects/x.md ). The agent attaches a document by reference instead of spending a tool round-trip to fetch it, and it can subscribe to that URI and get told the moment the note changes. Reading through a resource costs no tool definition. Reading through a tool costs one on every request. Use the right primitive. 3. Fail closed on auth, fail open on limits A token with no grants

Despite what influencers may say, you don’t need to spend $99.99 on Dyson’s HushJet Mini Cool or $149.99 for the Shark ChillPill to survive the summer whenever you leave the comfort of air-conditioning. My family has found all the comfort it needs to survive humid baseball games, sweltering concerts, and sweaty hikes with a couple […]

Following the original's debut at CES earlier this year, Twelve South is introducing a new version of its leather-wrapped Valet charging tray designed for use in places where space is at a premium. While the original Valet was 7.5-inches deep and large enough to serve as a catchall for a couple of items like your […]

Getting a research agent to call three tools in a demo is easy. The hard part starts when the seventh tool call times out and the first six have already spent money and changed state somewhere. So this is a question about recovery, not about AI frameworks. Temporal and Diagrid Catalyst both do durable execution, and both position themselves for AI workloads. What separates them is what each one asks you to build and operate around your agent. Start with the failure contract Say your agent searches internal documents, calls an external research API, asks an LLM to synthesize the evidence, and then waits for a human to approve the result. Before that design goes to production, you need answers to four questions: Which completed steps will not run again after a crash? How do you stop non-idempotent tool calls from firing twice? Can the agent wait for hours without holding a process open? Can an operator reconstruct what the agent did, and why? Temporal treats failure-prone work as activities, coordinated by durable workflows. It persists workflow state and rebuilds it by replaying history. You can self-host Temporal or use Temporal Cloud, and Temporal's current materials cover agentic applications and framework integrations directly. Catalyst builds on Dapr Workflows. Your agent runs as a durable workflow, and Catalyst ships runners for established agent frameworks. Diagrid's documentation describes a shared runtime layer that handles durability, workload identity, policy enforcement, and operational visibility across agents, workflows, MCP servers, and applications. Where the approaches diverge Temporal fits when you want your application logic written in its workflow-and-activity model. The conceptual model is mature, language support is broad, and there is a large body of distributed-systems guidance to learn from. The cost is fluency. Someone on your team has to understand Temporal's execution semantics well enough to reason about replay, and you still decide separ

After a lovely and productive conversation with your client, with still ringing ears, you check the coding agent's last log messages on a ticket that adds a discount to a product. The message was: "Done, I added the 10% discount and all tests pass. Stopping. " Well ... you know it's just not true, so you dig further and quickly realize that the discount functionality was never actually added and the tests it reported passing had never been run. The agent reached the end of the loop, looked at its own work, and called it finished. That call is the thing that shipped. This has a name. A paper published this June, From Confident Closing to Silent Failure , calls it false success: the agent asserts the task is complete while the actual state of the system says otherwise. It is common, and it holds up across capable models. On AppWorld, a benchmark for long-horizon coding agents, 75.8% of the runs that actually failed still ended with the agent claiming it was done. The researchers then put five different LLM judges on those completion claims, varying the prompts each time, and every one of them landed barely above a coin flip, because the thing each judge was reading was the closing sentence, and the closing sentence reads as confident whether the work happened or not. What told a real done apart from a false one turned out to be cheap and mechanical: a look at the actual state of the system. A lightweight deterministic state check caught four to eight times more false successes than the best of the judges. The paper has a name for the mechanism underneath, a hallucination of verification: the model narrates having checked something it never checked, and that narration is indistinguishable, sentence for sentence, from a report of a check that really ran. That gap, between what the agent said and what the system did, is what this piece is about. A loop runs five arms: generate, check, steer, retry, stop. The series opener named them; four pieces since took the check that

If you've been following this series, you know I've been testing BrowserAct for months now. Article 1 covered the CLI setup. Article 2 covered headless + human handoff. Article 3 was a 6-week production review. Those were all about the CLI, the developer tool. This article is different. BrowserAct now has a cloud product called BrowserAct Agent Built where you describe what data you need, and it builds a reusable scraper for you. No terminal. No code. Just a prompt. I tested it on five real business workflows. Here's what I found. Every quarter I update a pricing comparison spreadsheet for my clients. I work with teams evaluating deployment platforms, and the question is always the same: "Which one should we use for this project?" The honest answer depends on workload, team size, and budget. So I maintain a comparison across Vercel, Netlify, Railway, Render, Fly.io, and DigitalOcean. Six platforms. Six tabs. Two hours of squinting at marketing copy and copying numbers into a sheet. I wrote Python scrapers to automate it. BeautifulSoup, Playwright, the works. They lasted three months. Then Vercel redesigned their pricing page. Selectors broke. Fixed them. Netlify changed theirs two weeks later. Fixed again. Fourth breakage in six months, I stopped maintaining the scripts entirely. Back to manual. Two hours, every quarter. For a spreadsheet. But here's the thing: across my client engagements, I keep seeing the same problem in different shapes. The e-commerce team tracking competitor prices on Amazon every Monday. The agency paying for lead lists that are already stale. The HR team spending days copy-pasting salary data from job boards. Everyone needs web data. Almost nobody wants to maintain the code that collects it. Yesterday I tested BrowserAct Agent Built on five business workflows I actually deal with across different client engagements. One prompt each. No code. No selectors. Results below. Table of Contents What BrowserAct Agent Built Is (Quick Context) Test 1:

Apple's new program gives you lease options for iPhone, iPad, Apple Watch and Mac.

Cómo montar un motor de contenidos que no te arruine (julio 2026) Si sigues pagando 300 euros al mes por herramientas "todo en uno" de marketing, estás tirando el dinero. A mediados de 2026, la tecnología para automatizar ha bajado tanto de precio que los costes de infraestructura de contenidos son casi ridículos. La clave no es la herramienta cara, es conectar piezas pequeñas con APIs baratas. Aquí tienes cómo tengo montado mi flujo de trabajo ahora mismo. La pila tecnológica (el stack) Para automatizar sin gastar, olvida las plataformas de marketing tipo HubSpot o plataformas cerradas. Mi setup actual es este: Cerebro: Claude 3.5 Sonnet (vía API). Es mejor razonando que GPT-4o para tono editorial. Orquestador: n8n (corriendo en una VPS de 5 euros al mes en Hetzner). Base de datos: Notion (vía API para gestionar el calendario). Distribución: Ghost para el blog y la API de LinkedIn/X para el alcance. Coste total: Menos de 15 euros al mes. Paso 1: El disparador (el calendario en Notion) No uses un Excel. Usa una base de datos de Notion con cuatro columnas: Estado , Título , Prompt_Contexto y Fecha_Publicación . Cuando cambias el estado de "Borrador" a "Listo para generar", el webhook de n8n se dispara. Aquí es donde empieza el ahorro. No envías toda la base de datos, envías solo el registro nuevo. Paso 2: El prompt como código, no como texto La mayoría de la gente comete el error de pedirle a la IA: "escribe un post sobre X". Sale basura genérica. En 2026, si no das contexto, el contenido no posiciona ni recibe interacción. En tu nodo de n8n, construye el prompt de forma dinámica. Así es como envío la estructura a la API: { "model" : "claude-3-5-sonnet-20260620" , "system" : "Eres un redactor técnico senior especializado en SaaS B2B. Tu estilo es directo, sin paja, sin adjetivos innecesarios. Evitas los clichés de marketing de 2024. Tu objetivo es educar, no vender." , "messages" : [ { "role" : "user" , "content" : "Escribe un artículo corto basado en este punto clav

Our production chain is deterministic end to end: same configuration, same build, same app. That's what a platform promises. And a few months ago, we plugged into it the least deterministic component in existence: a language model, tasked with producing code that will be installed in a customer's app. What happens to that code — how it fits in, what it's allowed to touch — is a story I've told elsewhere . This one is about the machinery around the model: what you have to build so that a component that never answers the same way twice can live inside a chain that isn't allowed to vary. The output isn't an answer — it's a proposal That's the first mental shift. When the model hands back its work, nothing treats it as a result: it's a proposal — and it's about to be inspected. The rules are non-negotiable: one root container; no script sneaked into the HTML — external dependencies are declared in the open; HTTPS everywhere; only approved domains are reachable. None of this is politely requested in the prompt in the hope the model remembers. The prompt educates; the validator decides. Everything gets re-checked mechanically, after the fact, on every generation. That's the difference between trusting a component and putting it under contract: you don't ask it to be reliable — you make its unreliability harmless. The loop has a hierarchy When validation rejects a proposal, we don't call the model back right away. The response is tiered, cheapest first — because in production, every model call costs time and money, and a deterministic program that knows how to repair beats a regeneration that might. At the bottom: mechanical repairs. Malformed JSON gets fixed without a model. A missing field in a revision gets compared against the parent: if everything else matches byte for byte, the missing field is inherited — the model had simply judged it unchanged, and it was right. And when one block is genuinely missing, we make a micro-call for that block, not for the whole section