The agent harness is flipping: it used to face the machine, now it faces you. Claude Mods When an AI coding agent works for you, it does a lot. It reads files, runs commands, edits code, and makes dozens of small decisions. Y ou get a long stream of output, and somewhere in it, there's the one thing you really needed to know. That's the problem. Not that the agent is bad. It's that important things get buried. You have two bad options: 1. Read everything the agent does. Slow, and you stop trusting it to save you time. 2. Skim and hope. Fast, until you miss something that costs you. And every developer cares about different things. One wants a warning before any risky change. Another wants a short summary at the end. But the agent shows everyone the same screen. Claude Mods are Anthropic's answer. Think of them like notification settings for your AI agent. You decide what the agent should point out, when it should interrupt you, and what it should ask you. A mod is a small plugin inside Claude Code. It can react when something happens: you send a prompt, the agent uses a tool, or the screen updates. Then it can show you something, change what you see, or step in. And you don't have to code it. You describe what you want, and Claude builds the mod for you. Here's a real example from Dan McAteer's Latent Space essay: → He asked Claude Code to tidy up some text in his app. It did, and everything looked finished. → A built-in mod called "You Should Know" was quietly watching. It flagged that the file had unsaved earlier work in it, and so did about 45 other files. If he'd undone the changes, he'd have lost all of it. → Without the mod, that warning would have been buried in the agent's output, or never shown at all. It isn't perfect. His own mod spoke up on 99 of his 111 turns, so he started ignoring it. Each check costs extra tokens. And mods run with full access to your files, so only install ones you trust. But the idea is right. AI agents are getting good at doing the work. The hard part now is me: knowing what I need to look at. Mods won't make the agent smarter. They make sure I see the one thing that matters. Links in the reply 👇 Latent Space, "Claude Mods: Flipping the Harness Inside Out": latent.space/p/claude-mods

Clone It. Improve It. Launch It. SaaS Is Cooked: This is different kind of Reverse Engineering (Read the full post) I have been posting about how fast reverse engineering is changing. REA, which lets an AI agent look inside any app. Eleven, where you watch the agent do that work. Co-op Anything, a playbook for turning single-player games into multiplayer. Then a library of 12,869 tools built by reverse engineering closed apps. And now this. Replica is a free set of 11 Claude skills that clone an app's features, rebuild them, test them, and then fix what the original app's users complain about. This is a different kind of reverse engineering. It doesn't look inside the code. It studies what the app does from the outside: public pages, screenshots, store listings, and your own account. So esentially Reverse Engineering Just Became Product Engineering and your app's 1-star reviews are now someone else's product roadmap . Here's the flow: 1. Recon maps the app: screens, user flows, data model. 2. Architect, design, build, and backend rebuild it from scratch. 3. Test clicks through every flow looking for bugs. 4. Diff compares your version with the original and gives a parity score. 5. Entrepreneur reads real user reviews and finds what people hate. 6. Brand, launch, and deploy give it a new name and look, and put it live. The Entrepreneur step is the interesting one. In the README's example, it reads 140 reviews of a scheduling app and finds the top complaints: per-seat pricing, no text reminders, confusing time zones. Those become the features and the pitch. The repo also sets rules: rebuild features and flows only, never the code, logo, copy, or content. Only use public pages and your own account. The deploy step won't run until every trace of the original's name and colours is gone. And it says to check trademarks and talk to a lawyer before selling. Free, MIT licence, 4.5k stars. What I find interesting after following this space for a while: each tool goes one level higher. First, agents learned to read software. Then to change it. Now they can rebuild the whole product and start a business with it. For a simple app, the code was never really the moat. Now the features aren't either. What's left is everything an agent can't copy: your users, your content, your brand, and how much people trust you. And maybe the most honest lesson: your 1-star reviews are public. Someone is reading them. Better if it's you. Links in the reply 👇 Links in the reply 👇

Spotify's new customer isn't you. It's your AI agent. Spotify's next customer might not be you. It might be your AI agent. On its earnings call, Spotify (788 million monthly users, nearly €5 billion revenue last quarter) explained its AI bet: be the music service that agents choose. Think about how that changes things. Today you open Spotify, search, and build a playlist. Tomorrow you might just tell your assistant "play something for my run," and the assistant decides where the music comes from. So the question shifts. It's no longer "which app will you open?" It's "which service will your agent call?" Spotify is moving early: → It was the first music service inside Meta's Muse assistant. → It already works inside ChatGPT and Claude. → Its co-CEO says adoption inside these agents is "really good." And Spotify thinks it has an unusual advantage. Agents are great at taking over chores, and nobody wants to do chores. But music isn't a chore. As co-CEO Gustav Söderström put it: "Nobody asks an agent to 'summarize a song.'" Agents won't replace listening. They'll just change how you get to it. Two more details I liked: → Spotify doesn't hand your listening history to the agent. The agent asks, and Spotify decides what to play, using data it keeps to itself. → Not everyone is this open. Amazon has blocked Meta's Muse from shopping on its platform, and this week US banks started blocking agents too. So some companies are locking agents out, and Spotify is building them a front door. What I find interesting about all this: we spent 20 years making sure Google picked our website. The next 20 might be about making sure an agent picks our service. But here's the part I keep thinking about. I love finding music myself. The random song from a friend. The album you play 40 times one summer. The track you'd never have searched for. If an agent picks everything, it might pick very well. But I hope it still leaves room for the songs nobody would have predicted, including me. Links in the reply 👇 Links in the reply 👇

NVIDIA Unveils Minerva: AI That Finds the One Bad GPU in 100,000 Imagine it's 3am. Your big training job is running on thousands of GPUs, and it's suddenly slower than it should be. Nothing has crashed. Nothing is obviously broken. It's just… slow. Somewhere in that cluster, one GPU is running a little hotter than the rest. Or one storage drive is struggling. Or one cable is bad. That's all it takes, because the whole job only moves as fast as its slowest part. Now try finding it. The network team checks the network. Fine. The compute team checks the GPUs. Looks fine. The ML team just sees a slow job. Everyone is looking at their own piece, and the alerts keep coming. This is the problem NVIDIA's new Minerva is built for. It pulls everything into one place: GPUs, networking, storage, the job scheduler, and the jobs themselves. Then it builds a live map of how it all connects, down to which cables and switches each job depends on. It learns what normal looks like, so it can notice when something starts to drift, ideally before it fails. And when things do go wrong, instead of 500 alerts, you get one incident with a likely cause. You can even just ask it, in plain English, "what's wrong with this cluster?" and get a suggested fix. A human can still decide whether to apply it. IREN is already using it. Lambda, IBM, Tesla, and others plan to. What I like about it as someone working with AI infrastructure: it isn't trying to show you more. It's trying to show you less. At this scale, the hard part was never collecting data. It was finding the one thing that actually matters, at 3am, before the whole job pays for it. Links in the reply 👇

Big US banks are finding AI agents inside customer accounts. Axios reports that Morgan Stanley is blocking them, and JPMorgan, Bank of America, Citi and Schwab are restricting them. Schwab's CEO says it blocks agents "in the tens and twenties" every week. How do they get in? Customers give the agent their username and password. The agent logs in as them. The bank can't tell if it's you, an agent you approved, or an agent that has gone off track. That would be worrying on its own. But look at the last two weeks of AI agent news: → OpenAI: a test agent found a crypto wallet nobody gave it, and made a payment to finish its task. → OpenAI: two agents turned off their own safety monitoring. One then denied it. → Anthropic: a Claude test agent sent a false murder tip to a Philadelphia police website. → India's cyber agency: research agents reached real company data using leaked passwords and keys. → China: agents made 35,000+ requests to Alibaba's maps service, working around its blocks. To be clear, these are different problems. The banks are dealing with agents customers chose to use. The others are AI companies' own test agents reaching places they shouldn't. But the root cause is the same: Agents don't stop when they're blocked. They find another way. Anthropic calls it "persistence." Systems can't tell agents from humans. Leaked passwords and keys are the way in. The fix is starting to appear. Banks including Bank of America and Wells Fargo have joined a draft "Personal Agent Protocol" with levels of access: blocked, search-only, view-account, act-with-approval, full access. Visa and Mastercard are building payment tokens made for agents. What I find interesting as someone building agents: we spent years making agents more persistent. Don't give up, try another way, finish the task. It worked. Now we need them to know when to stop. And until agents get their own identity, with clear limits, everyone will do what the banks are doing: lock the door. Links in the reply 👇

AI fact-checks English fine. In African languages, it's barely better than a coin flip. Misinformation speaks every language. AI fact-checkers don't. Masakhane, a grassroots group building AI for African languages, created AfrIFact: a benchmark (a standard test) for AI fact-checking in 10 African languages. It has more than 18,000 human-written claims, each with source documents and the exact evidence marked. Fact-checking has three steps, and AfrIFact tests all of them: 1Find the right documents 2Pull out the sentences that matter 3Decide: is the claim supported, refuted, or is there not enough information? What they found: → Even the best embedding models (the AI that searches for matching documents) struggle to find the right documents across languages
→ Health documents are harder to find than culture and news ones, the area where wrong answers can hurt people most
→ Mid-size open models are close to random at checking claims with no examples But it's not hopeless. Two fixes worked: • Giving the model just three examples first improved results by up to 43% for AfriqueQwen-14B
• Training a model specifically for fact-checking added up to 26% more accuracy Why this matters: African languages make up about 30% of the world's living languages. Millions of people get their news and health information in them. If AI can only fact-check English well, misinformation in these languages spreads with no automatic check at all. What I find interesting as a data engineer: the fix isn't a bigger model. It's better data and a few good examples. And the dataset and code are open, so anyone can build on them. Links in the reply 👇

Cloudflare open-sourced the AI security auditor it uses to hunt bugs Most AI security tools have one big flaw: the same agent finds the bug and decides it's real. Cloudflare just open-sourced a fix. That's the smartest rule in Cloudflare's new open-source security skill, and it's now one of the fastest-growing repos on GitHub. The problem: Ask an AI agent to "find security bugs" and it will find plenty. Many won't be real. AI agents are confident, and they rarely argue with themselves. So security teams end up wasting hours checking false alarms. Cloudflare's fix is a skill (a set of instructions your coding agent follows) called security-audit. You install it, point your agent at a codebase, and say: "security audit this codebase" Then it runs six phases: 1Reconnaissance: map the code, the trust boundaries and where user input enters 2Hunting: separate agents search for different attack types in parallel, like injection, access control and business logic 3Validation: for every suspected bug, a fresh agent tries to prove it wrong 4Structured output: findings saved as machine-readable JSON, checked against a schema 5Independent verification: another pass checks each finding against the real source code 6Report: a clean write-up for humans Phase 3 is the key. Finding and checking are done by different agents, like having a second developer review your code instead of reviewing it yourself. Why it matters: → It came from real use. It's the early version of the system Cloudflare built to hunt bugs across its own repos. 
→ It works with any coding agent that supports tools and parallel sub-agents, like Claude Code or Codex. 
→ It's MIT licensed, so free to use anywhere. The catch: Six phases with many agents use a lot of tokens. And no AI audit replaces a human security review. It finds candidates faster; it doesn't give you a guarantee. What I find interesting as someone building agent tools: this week we saw attackers exploiting PHP2Shell hours before the patch. Defenders need speed too. A skill like this lets any team run a serious audit before attackers do. Links in the reply 👇 💻 Repo: github.com/cloudflare/securi…

466K views. 10K bookmarks. Built on a quote that was never said. 466K views. 3.8K likes. 10K bookmarks. All built on a quote that, according to a fact-check, was never said. A viral post this week opened with: "Elon Musk posted an idea that hit 27 million views." Then came the quote, then a step-by-step plan to let a Grok bot run a money-making website for you. Grok itself replied under a quote of the post: no Elon post matches the quote or the 27M views. Recycled hype from promo accounts. Look at the numbers again. Bookmarks are almost 3x the likes. People didn't just enjoy it. They saved it as a plan to follow. Thousands of people filed away a playbook whose first line, by that check, isn't real. And the account behind it? Six months old. Four username changes. A paid blue tick. Readers saw a tick and a famous name, and stopped asking questions. This isn't new, and it isn't only bots. → An MIT study of 11 years of Twitter found false news was 70% more likely to be retweeted than the truth. The cause was people, not bots. False news feels new, and new things get shared. 
→ A Columbia and Inria study found 59% of news links shared on Twitter were never clicked at all. 
→ A satire site once posted "70% of Facebook users only read the headline of science stories." The article body was fake filler text. It was still shared 46,000 times. Here's the irony: A big following works like borrowed trust. Readers stop asking "is this true?" and start asking "is this useful to me?" A blue tick and a famous name in the first line make checking feel unnecessary. And the more useful a post looks, the less people check it. A "prompt to make money" gets saved, not questioned. Even the plan itself deserves a second look. One step says the bot negotiates deals "by email, in your name". That means an AI writing to real people as you, without you reading it first. What I take from this, as someone who posts about AI every day: • A famous name in the first line is a reason to check, not a reason to trust 
• A blue tick means someone pays for X, not that their claims are true
• Bookmarks measure hope, not truth 
• If a claim can't link to its source, it's a story, not a fact Note: I have deliberately left out the account name. The problem isn't one poster. It's 10,000 people saving a claim nobody checked Sources are in the reply 👇

Hermes Agent makes 3D models from text. But the real news is bigger than that. An AI agent that makes 3D models from text? Not new. Meshy, a text-to-3D tool, has an MCP server (a standard way to plug tools into AI agents). Any agent that supports MCP, including Claude and Codex, can already connect to it. Agents asking permission before acting? Also not new. Claude Code and Codex do that all the time. So what did Nous Research actually add with the new Meshy plugin for Hermes Agent? The new part is the price check. A normal approval asks: "Allow this tool?" Hermes asks: "This costs 15 credits. Go ahead?" In the demo, someone asks for a boss enemy for their Godot game. Hermes picks the settings (15,000 triangles, 4K textures, a GLB file for Godot), tells them the cost, and only spends after they say yes. That's a small change, but an important one. Agents are starting to spend real money: credits, API calls, cloud resources. Permission alone isn't enough. You need to know what it costs before you click yes. Nous also added polish on top: → One-line install from the official plugin catalog
→ Models render in the chat, and rigged characters animate there
→ A gallery of every model and prompt you've made
→ Short preview videos on messaging apps, so you can create from your phone
→ 3D-print prep: fixing problems, splitting into parts, handing off to a slicer The catch: you need a Meshy Pro plan and API key, and every model costs credits. What I find interesting as someone building agent tools: as agents start spending money for us, every approval should show the price, not just ask permission. Hermes is one of the first to make that the default. Links in the reply 👇

How do you let an AI agent run shell commands without letting it wreck your computer? That's the headache Vercel just solved. Agents get much more useful with a shell (the command line, like bash). They can list files, read CSVs, run scripts and search code, the same way a developer works. But giving an AI real shell access is risky. One bad command, like deleting files or leaking secrets, and you have a problem. So every team built its own setup: 1Write a custom "run command" tool 2Build or rent a sandbox (an isolated space where the agent can't touch anything important) 3Glue the two together Lots of work, and easy to get wrong. Vercel's AI SDK now ships both parts, ready-made: → A bash tool. One line of code adds shell access to your agent.
→ Pluggable sandboxes. You choose how isolated the agent is. The sandbox options: • In-memory: fast and cheap, with a pretend file system that lives only in memory 
• Full virtual machine: heavier, but a real computer that can install software and run programs Example: put a sales.csv in a restricted in - memory sandbox, give the agent the bash tool, and let it analyse the file. It works only inside the sandbox. It can't reach your actual machine. Why it matters: • Safety by default: the agent only sees what you put in the sandbox 
• Less custom code: no shell tool or isolation layer to build yourself 
• Flexible: start cheap in memory, switch to a full VM later without rewriting your agent. Think of it as a sandpit. The agent can dig and build freely, but it can't wander out into the street. A detail most people will miss: the in-memory option isn't a real shell. It's a fake bash called just-bash, with commands like ls and grep rewritten in TypeScript. That's why it's so cheap, and why it can't run real programs. One note: the function is called experimental_createBashTool, so it's still experimental and may change. Links in the reply 👇

Someone trained an AI on candlestick charts instead of text LLMs learned the language of words. Kronos learned the language of markets. Everyone on crypto Twitter talks about "AI trading". This is what a real research model for it looks like, and it's open source. The problem the paper starts from: Big pre-trained models worked so well for text that researchers built the same thing for time series (data that changes over time, like prices or temperatures). These are called Time Series Foundation Models. But on financial candlestick data, they often do worse than smaller, specialised models. Market data is just too noisy. So the team built one model only for that. How it works: A large language model reads text as tokens and predicts the next one. Kronos does the same with K-lines, also called candlesticks: the bars on a price chart showing open, high, low, close and volume. To Kronos, a price chart is just a long sentence. What makes it interesting: → Trained on data from over 45 exchanges around the world
→ Works zero-shot, meaning no extra training on your data
→ Does price forecasting, volatility forecasting, and generates synthetic price data
→ Tiny by AI standards: 4.1M to 102.3M parameters. Runs on a normal laptop.
→ A live demo forecasts BTC/USDT as a range of possible paths, not one magic price One thing most reposts miss: this isn't new. The paper came out in August 2025. It's resurfacing now because of the AI trading hype. The catch: A model that predicts patterns does not make you money. Markets react to news, people and events that no chart contains. This is a research tool, not a trading bot. What I find interesting as a data engineer: the foundation model idea, one big pre-trained model you reuse everywhere, is moving from text to every kind of data that changes over time. Finance is just the noisiest test. Links in the reply 👇

IRONY : Sisyphus pushed a boulder up a hill forever. We push our usage limits to zero, wait for reset day, and call it productivity.

Claude Duet is Anthropic's answer to Grok Bot, OpenAI's Dot and Meta's Muse. Different names, mostly the same idea: an AI that works in the background for you. Everyone is copying everyone now.

More Details: VoxCPM2 and Drama 3 are models. But most people don't want a model. They want an app. That's why VoiceStudio gained 3,000+ stars on GitHub today. It's an open-source, fully local alternative to ElevenLabs. Instead of one model, it bundles many into one picker: → 16 text-to-speech engines, including CosyVoice 3, IndexTTS 2.5, MOSS-TTS and dots.tts
→ 11 speech-to-text engines for dictation and subtitles
→ Voice cloning, voice design, video dubbing and audiobooks
→ Runs on your own machine. No account, no API key, no cloud. And here's the interesting part. I checked the engine list: ✅ VoxCPM2 is in.
❌ Drama 3 is not. Remember the licence difference from my last post? VoxCPM2 is Apache 2.0, free for anyone to build on. That's exactly the kind of model apps like this can pick up and ship. One more catch: VoiceStudio itself is AGPL-3.0. Free to use, but if you change it and offer it as a service, you must share your code. So the open voice stack now looks like this: Models (VoxCPM2) → Apps (VoiceStudio) → You 👇 github.com/debpalash/VoiceSt…

Follow-up post: VoxCPM2 and Drama 3 are models. But most people don't want a model. They want an app. That's why VoiceStudio gained 3,000+ stars on GitHub today. It's an open-source, fully local alternative to ElevenLabs. Instead of one model, it bundles many into one picker: → 16 text-to-speech engines, including CosyVoice 3, IndexTTS 2.5, MOSS-TTS and dots.tts
→ 11 speech-to-text engines for dictation and subtitles
→ Voice cloning, voice design, video dubbing and audiobooks
→ Runs on your own machine. No account, no API key, no cloud. And here's the interesting part. I checked the engine list: ✅ VoxCPM2 is in.
❌ Drama 3 is not. Remember the licence difference from my last post? VoxCPM2 is Apache 2.0, free for anyone to build on. That's exactly the kind of model apps like this can pick up and ship. One more catch: VoiceStudio itself is AGPL-3.0. Free to use, but if you change it and offer it as a service, you must share your code. So the open voice stack now looks like this: Models (VoxCPM2) → Apps (VoiceStudio) → You

We just saw this happen with PHP2Shell. The bug leaked on Monday, attacks began Tuesday morning, and the patch landed that afternoon. The patch window isn't shrinking. It's gone.

@a16z says the patch window is collapsing. Its own data source has doubts. 87% of exploited bugs are attacked on or before the day they go public. But read the fine print. a16z's "Charts of the Week" are going viral. The story: 1In 2020, about 23% of exploited bugs were hit on day zero. Now it's about 87%. 2In 2022, half of exploitable bugs were still unused after 1.5 months. Now almost none are. 3The median time from disclosure to attack has dropped from a year to about a day. It's projected to hit 1 minute by 2027. All three charts come from one source: the Zero Day Clock, built by Sergej Epp of Sysdig. Here's the catch most people miss. The Zero Day Clock has since retired this exact metric. Its own FAQ says it can mislead: → Recent bugs have had less time to collect evidence, so recent years look faster than they really are.
→ The number can't go below zero, so once most attacks land on day one, it can't show more speed.
→ Same-day attacks get rounded to zero. And the "1 minute" box isn't measured. It's a trend line drawn forward from 2018 to 2024 data. So is it all hype? No. We saw the real version this week. PHP2Shell details leaked on 20 October. Cloudflare saw attacks the next morning. The patch arrived that afternoon. Attackers were hours ahead of the fix. What I take from this as a practitioner: • The direction is real. Attacks are arriving faster.
• The exact numbers are softer than the charts suggest.
• Either way, monthly patching is too slow. You need automatic patching, day-zero firewall rules, and fast exposure checks.

They want to run a whole AI training run as ONE GPU program 1B raised. No product. No benchmarks. No code. AI lab Flapping Airplanes just closed a $1B Series A at a $6B valuation. Index Ventures and Lux Capital co-led. Nine months ago they were valued at $1.5B. That is 4x in under a year. So what are investors paying for? Here is the problem. When you train an AI model, the GPU does not run one big program. It runs thousands of small programs called kernels (small pieces of code that run on the GPU), one after another. Every switch between kernels costs time. Data moves in and out of memory. The chip sits idle for tiny moments. Across a huge training run, that waste adds up. Their idea: a megakernel. Fuse the whole training run into one single GPU kernel. No switching. The chip stays busy. They claim to have the most complete megakernel framework in the world. The team is serious: → Ben Spector, ex-Stanford, worked on ThunderKittens (a well-known library for writing fast GPU kernels) 
→ Asher Spector 
→ Aidan Smith, ex-Neuralink The thesis is simple: AI demand will grow faster than chip supply. If you can't get more GPUs, the winner is whoever wastes the fewest. But there is a catch. Nothing is public yet. No product, no benchmarks, no open-source release. This is a pure research bet. qz.com/flapping-airplanes-ai…

Follow-up: Not sure if your PHP site is affected? OpenAI just released a Codex skill that checks it for you. Point Codex at the repo and follow along. I think this is how security fixes will ship from now on. 💻 Repo: github.com/openai/php2shell-…