1B Runs Logged on Weights & Biases, CoreWeave ARIA Integration

TL;DR: The new auto mode announcement just confirms what we landed on last time about taste. Part 3 of my (kinda) series about agentic engineering and what it means for your skills. When fellow engineers ask me how I configure my Claude Code sessions, I usually answer "auto mode only, anything else is just pretending". Once the plan stands and you agree with it, the code writing part is the boring part. And be honest with yourself here: how often do you really read AND understand what Claude is asking you to approve? Anthropic just announced auto mode becomes the default on August 14, and the reasoning they published is the interesting bit. In a study with 1,053 paying testers they put a clearly dangerous command into a permission prompt. People caught it 13.6% of the time. After 50 prompts that dropped to around 5%. The classifier caught 89%. So clicking approve on prompt number 40 has very little to do with judgment, and I think most of us knew that already. It's muscle memory with extra steps. To be fair, Anthropic is careful about this themselves. They point out that the classifier reduces risk without eliminating it, and they still recommend keeping a human in the loop for high-stakes changes to production infrastructure. I agree with that, and I'd add that it only works if you actually look when it matters, which is hard to do when you've spent all day rubber-stamping prompts. Which brings me back to where I landed a few weeks ago. The judgment I actually want to protect sits further up: is this the right thing to build, does it solve the problem we think it solves, is the idea any good in the first place. A classifier doesn't touch that. And approving 200 prompts a day never sharpened it either. How do you run your sessions? #EngineeringLeadership #AI #AgenticEngineering

My Journey at Pi Dot – #Part 7 Building is exciting. Testing teaches you the most. One thing I learned during this journey is that no matter how much time you spend building something, there will always be things you don't notice until you start testing it. Before every simulation, I spent a lot of time testing the platform from a student's perspective. I checked whether the AI personas responded according to their roles, whether the workflows were working as expected, and whether the overall experience felt smooth from start to finish. Many of the improvements we made came from these testing sessions. Sometimes it was a small UI issue, sometimes it was an AI response that didn't fit the business scenario, and sometimes it was a workflow that needed a small change. These weren't big changes individually, but together they made the platform much more reliable for students. This experience changed the way I look at products. Building a feature is important, but testing it, improving it, and listening to feedback is what makes it useful. In my next post, I'll share how every new simulation helped us improve the platform even further.

The world doesn't need another AI demo It needs engineers who can ship! 👇 That's exactly why we created Grokking Agents in Production, the course I built alongside Antonio Zarauz Moreno and Luis Serrano. Most people can get an agent working on their laptop. Very few know how to make it survive the real world: latency, cost, failures, monitoring, and all the messy edge cases that only show up once real users arrive. 👉 𝐓𝐡𝐢𝐬 𝐜𝐨𝐮𝐫𝐬𝐞 𝐜𝐥𝐨𝐬𝐞𝐬 𝐭𝐡𝐚𝐭 𝐠𝐚𝐩. It's for engineers who are done playing with toy examples and ready to build agents that actually hold up in production. Just the hard-won lessons and the real engineering it takes to ship. To everyone who completed it: you didn't just watch some videos. You put in the work, earned the certificate, and leveled up into territory most people never reach. 𝐏𝐫𝐨𝐮𝐝 𝐨𝐟 𝐞𝐯𝐞𝐫𝐲 𝐨𝐧𝐞 𝐨𝐟 𝐲𝐨𝐮! 🙌 The future belongs to those who can take AI from "cool demo" to "real system." Now go build 😎

Fable 5 is impressive. But that doesn't mean it should be your default for Claude. A pattern we're already seeing: People turn on the highest reasoning mode for everything. Or they don’t know that’s what they are using. Quick emails. Simple code. Basic brainstorming. That's like driving a Formula 1 car to the grocery store. A simple rule of thumb: ⚡ Quick chats, emails, brainstorming → Sonnet 5 💻 Everyday coding & analysis → Opus 4.8 🧠 Research, agents, finance & legal work → Fable 5 (High) 🚀 Extremely difficult coding → Fable 5 (XHigh) 🔥 Problems almost nothing else can solve → Fable 5 (Max) That's how you'll get more done—and spend less doing it. 👇 Have you put Fable 5 to work yet? What's the first real task you gave it, and how did it perform? #AI #ArtificialIntelligence #Anthropic #Claude #Fable5 #AIModels #GenerativeAI #Productivity #AIEngineering #AIAgents

I’ve decided to blog more about my stuff. Apparently, “my stuff” means machine learning, competitive ML, and a bit of running. New post: https://lnkd.in/epJu_jJ3 First up: I backed the wrong Numerai model again, cleaned up my backtesting and risk process, and ran a new half marathon PB. 🏃 🙌 Good intentions. Bad stakes. Hard lessons. #Numerai #MachineLearning #Running

When I’m learning a new model it’s great to be able to work on several experiments in multiple large variation sweeps per shot using Luma agent. It’s a great way to quickly build intuitions for the capabilities of that model. Luma | Dream Lab Julia Harris Jacob Russo Bryan Soegondo Tony Hou

Most of us spend a lot of time comparing LLMs. GPT vs. Claude vs. Gemini vs. Kimi. Here’s another way to think about it. Imagine every LLM is a horse. Some horses are stronger. Some are faster. Some have better endurance. Some have been trained for different terrain. Those differences matter, but only up to a point. Now imagine that everyone has access to the same horse. Why do some riders consistently travel farther, faster, and with fewer mistakes? Because of the harness. The harness determines how effectively the rider can guide the horse, communicate with it, and get the best out of its capabilities. That’s what tools like Claude Code, Cursor, Kiro, Windsurf, Codex, and others really are. They don’t replace the horse. They build a better harness around it. The best developer experiences aren’t just about which foundation model is underneath. They’re about the engineering wrapped around it: context management, memory, planning, tool orchestration, code editing, testing, terminal integration, guardrails, and workflow design. As foundation models continue to converge in capability, the competitive advantage will increasingly come from the harness rather than the horse. The horse provides the power. The harness determines how much of that power actually reaches the road.

This YouTube video stopped me mid-scroll. It argues we should stop reviewing AI-generated code line by line and start building constraint systems around agents instead. Automated test harnesses. Cross-model review. Back-pressure that catches bad code before it ships. I haven't tested it yet. But it's making me rethink how I validate agent output. Curious what people actually in the trenches think. Is this how you're working already? Or does it sound good in theory but break in practice? Link in the comments #AIEngineering #AgenticAI #claudecode #softwaredevelopment #buildinpublic

Have you ever noticed that when you are in the middle of your deepest, most focused work with Claude, that is when you get hit with usage limits? I had the same problem, so last week I tried to manage my credits better and figure out which model works best for each of my tasks. I think I managed well, only maxing out on weekly credits by Sunday evening, just in time for the World Cup final. Here’s how I now use each model: Fable: I have a rough idea of what I want but no idea how to build it. Or I need to clean up the work of other models. Opus: Most coding tasks and stuff that would normally take me hours to do. Sonnet: The daily workhorse for the smaller tasks and automations. Haiku: Not for me, it makes far too many mistakes. So if you are hitting your usage limits too quickly, try using different models next time and see if you can get what you need but still keep your credits intact. Remember, Fable will always be there to clean up the mess if it needs to. 😀 #Claude #ClaudeCode #AIProductivity #GenerativeAI

Heading to Las Vegas this week for Ai4 at The Venetian 🎲 A conference about non-deterministic systems, in a city built entirely on the fact that you can't predict the outcome! I can't wait to connect, learn, and share with AI leaders shaping what comes next. If you're there, come find me. I'd love to talk about how we're thinking about Context & Agent Trust at Monte Carlo. Here's the idea I keep coming back to: we spent decades perfecting the Software Development Life Cycle (SDLC), how about Agent Development Life Cycle (ADLC)? The two rhyme, but the differences are exactly where the hard problems (and the 2am pages) live. Learnings thus far: 1️⃣ Evals are the new QA Evals in agent development are the equivalent of QA testing in software. A great foundation. But like QA, you can't predict every way real users (and agents) will interact with your system, or the beautifully cursed questions they'll actually ask. Green evals and a live incident can absolutely coexist! 2️⃣ The Golden Dataset is controlled. Context is a group chat! Golden Dataset that sat still and behaved but Agents run on context: retrieved, dynamic, and changing while you're looking at it. You don't own that input surface anymore 3️⃣ Humans could observe and fix. Agents laugh at that plan. In traditional software, a human could observe, detect, and resolve issues. With agents, the complexity, dependencies, and sheer volume make manual oversight roughly as scalable as counting cards with your eyes closed. Trust has to be engineered in: observable, measurable, and continuous (Reinforcement Loops?) Software taught us that shipping is the beginning, not the end. Same for agents, except now the ground moves too. If this resonates, or you're quietly living it in your own stack, let's connect. I'll be at Ai4 all week, statistically somewhere near the coffee or at Monte Carlo Booth # 1260 👋 And if you like racing, join us Wednesday from 7 to 10 PM. #Ai4 #AIAgents #AIObservability #MonteCarlo #AgentDevelopment #DataTrust