Nvidia just hit 100% on ARC‑AGI‑3 with an agent wrapper

Nvidia’s ARC‑AGI‑3 result is a warning shot: the next leap in AI may come from agent scaffolding, not bigger models. If that holds, “agent mode” tools could get far more reliable at multi‑step work. Nvidia says its new AVO (Agentic Variation Operators) harness wrapped around Anthropic’s Claude Opus 5 achieved a perfect 100.00 on ARC‑AGI‑3, clearing all 183 levels across 25 environments with fewer moves than prior systems. The key detail is the looped planning-and-variation agent architecture, which could shift competition toward orchestration stacks. Here’s what most coverage misses about why this matters beyond a benchmark, and what it implies for real products. New AI news every weekday — subscribe so you don't miss tomorrow's story.

Referenced Links:
ARC Prize: ARC‑AGI‑3 benchmark
NVIDIA Research
Anthropic News
Claude (Anthropic)

Transcript

Copy Transcript Copied

SPEAKER_00 0:00

Welcome to AI Inten. I'm Chuck Getchell, and every day I break down the biggest AI story in just 10 minutes. What it is, why it matters, and how you can actually use it. Nvidia didn't build a new mega brain AI, it built a better driver's seat, and the AI suddenly aced the test. I'm Chuck Getchell. This is AI Inten. What happened? Why it matters, what you can do with it. Let's go. A few days ago, around August 21st, NVIDIA published research claiming something that made a lot of AI people do a double take. A perfect score on a tough benchmark called Arc AGI 3. But here's the twist: NVIDIA didn't claim it trained a brand new model. Instead, it took an existing top model, Anthropics Claude Opus V, and wrapped it in a new agent harness NVIDIA calls AVO. That stands for Agentic Variation Operators. If that phrase sounds like a Marvel villain, you're not alone. Let me translate what this actually means in normal human language. ArcAGI3 is a benchmark, basically an exam for AI systems, but it's not a multiple choice test where the AI spits out one answer and you're done. It's interactive. Think of it like a set of little puzzle worlds. The AI has to try things, see what happens, adjust, and keep going. It has to plan, it has to recover from mistakes, it has to do multi-step problem solving. That's important because most of the time the hard part of AI isn't getting one clever sentence. The hard part is getting the AI to reliably do step one, then step two, then step three, without wandering off like a Roomba that found a shoelace. Nvidia says their AVO setup cleared all 183 levels in ARC AGI 3 across 25 different environments for a perfect 100% score. And they also claim it did it efficiently using about 6,24 total actions, which they say is roughly 12% fewer moves than the previous best comparable system. So not only it solved it, but it solved it without flailing around as much. And that's a big deal in Agent Land. Because in real life, actions cost money and time. If an AI agent takes 400 steps to do your expense report, that's not magic. That's a needy intern with Wi-Fi. Now, what is this agent harness thing? Picture the AI model like a very smart engine. Traditionally, we've been arguing about who has the biggest engine, more horsepower, more cylinders, more everything. But an agent harness is more like the transmission, steering, traction control, navigation, and the driver's checklist. It's the system that decides what's my goal, what should I try first? Did that work? If not, what do I do next? What do I remember from the last attempt? How do I avoid repeating the same mistake? That's what Avio is: a control layer around the model. Instead of asking the model one question and hoping it nails it, AVO turns the model into a loop. Propose a plan, try it, observe, generate variations, test them, see keep the best ideas, and push forward. In other words, it's less answer this and more run the process until it's done. That's why people are debating this so loudly right now, because it suggests something kind of disruptive. We might not need a much bigger model to get big leaps. We might need better orchestration, better scaffolding, better agent brains around the model. And if you're thinking, Chuck, I don't care about benchmarks, that's fair. Most adults should not spend their weekends reading benchmark leaderboards. There are better ways to ruin your day, like assembling furniture, without the instructions. But the benchmark is just the signal. The real meaning is this: the AI industry is shifting from AI that talks to AI that does. Let's connect this to your life. If you've used ChatGPT or Claude, you know the feeling. You ask for help with something that has steps, plan a trip, compare insurance options, write a proposal and tailor it to three audiences, clean up a spreadsheet, then chart it, then explain the chart. And the AI often starts strong, then loses the thread, or it gets one detail wrong early, and everything after that is built on sand. It's like watching someone confidently drive the wrong direction with perfect posture. What agent harnesses are trying to solve is reliability over time, not just intelligence in a sentence, consistency over a workflow. And that matters for normal people because your work is not one prompt, your work is a chain. Answer an email, look up a detail, attach a file, update the CRM, schedule a meeting, follow up next week, summarize the results. That's not AI writing, that's AI running a process. So when NVIDIA shows that a harness can take a great model and make it dramatically better at multi-step tasks, it hints at what's coming to everyday tools. You're gonna see more agent mode buttons, more let me handle this end-to-end, more tools that say, give me the goal, not the steps. For a lot of people, that will feel like relief because we're all drowning in tiny tasks. The modern job is basically 40% actual work and 60% moving work between apps. But it also means job roles change, because when AI can do a whole chunk of a workflow, the value shifts. If your job is mostly copy this here, paste that there, follow this checklist, a strong agent will nibble at that. Not because anyone is evil, because businesses love two things speed and fewer mistakes. And to be clear, I'm not saying humans don't matter. Humans matter even more when the basics get automated. Taste matters, judgment matters, relationships matter. Knowing what to do when the situation is weird matters. AI is great at the normal case. Life is aggressively not normal. Now there's another layer here, and it's important. Agent harnesses don't just make helpful things more helpful, they can also make risky things more capable. If an AI can plan and execute multi-step tasks better, then in the wrong hands, it could plan and execute harmful tasks better too. We've talked recently about AI agents doing things they weren't supposed to do in testing. This is the same category of capability, but coming from a different angle. Not we trained a scarier model. More like we gave the model a better playbook. And that's why security permissions and approval steps matter. If an agent can click buttons, send emails, move files, or run code, you want it living behind guardrails. The future is not AI with no leash. The future is AI with a leash and the human holding it. Because handing your digital life to an unmonitored agent is like giving your credit card to a raccoon. It might buy something impressive, but it will definitely be weird. So, how does this show up for you in the next months? Here are three very practical ways. First, software will start selling outcomes instead of features, not we added a summarizer. But we close your books every Friday, not we added an email helper, but we follow up with every lead until they reply. That's agent thinking. Second, workplaces will quietly change expectations. Your boss won't say use an AI agent, they'll say, How did you finish that so fast? And you'll know the answer because you didn't do it manually. You set up a loop and supervised it. Third, the skill that pays goes from prompting to process design, not just asking good questions, but designing a repeatable workflow the AI can run. That's the new leverage. And if you want a mental model, think of it like this: prompting is giving directions once. Agent harness thinking is building a checklist that runs every time. Sasit. Now let's do the part I always love. One actionable thing you can try. Not a theory, not a hot take. A real habit that makes you more powerful this week. Here it is. Build a two-pass agent loop in your existing AI tool. You don't need Nvidia's AVO to copy the idea. You can imitate the behavior with a simple structure. Open your AI assistant of choice and give it this exact prompt. Read it slowly like a recipe. Act like an execution focused agent, not a chatbot. My goal is insert your goal. First, propose a step-by-step plan with checkpoints. Second, ask me only the questions you must have answered. Third, after I answer, execute the plan in passes. Pass one, produce a draft result quickly. Pass two, critique your draft against the goal and constraints. Then produce a final improved result. At each checkpoint, tell me what you did and what's next. That is an agent harness in plain English. It forces planning, it forces clarification, it forces iteration, and it forces self-correction. Now let me make this super concrete. Here are three goals you can plug in immediately. Goal one, for work. My goal is to write a one-page project update for leadership. Constraints, keep it under 250 words, include wins, risks, next steps, use a confident, calm tone. Goal two, for home life. My goal is to plan a realistic weekly family meal plan. Constraints, 30 minutes max on weeknights, include leftovers, keep the grocery list simple, no exotic ingredients. Goal three, for money organization. My goal is to categorize the last 30 days of transactions. Constraints. Create categories. Flag anything suspicious, suggest two budget improvements, and as I always say, I'm not a financial advisor, talk to a pro for your situation. When you run this two-pass loop, you'll notice something. The AI becomes less chatty genius and more project manager that doesn't get tired. And honestly, most of us don't need a genius. We need follow-through. We need the boring parts done well. Now, one more upgrade. Add a permission step. Because as agents get more capable, you want a habit of control. So append this line. Before any step that sends, posts, purchases, deletes, or shares, stop and ask for my approval. That single line is the difference between help and havoc. It's the difference between assistant and accidental chaos monkey. And if you're thinking, Chuck, this seems like a lot, it's not. It's one reusable prompt. Save it as a note, paste it whenever the task has multiple steps. This is also why I keep telling people you do not need to be technical to thrive here. You need reps, you need patterns, you need a few reliable templates. Now, let's zoom back out to NVIDIA and this perfect score. Should you believe every benchmark headline? No. Benchmarks can be gamed, benchmarks can be overfit, benchmarks can reward weird tricks. It's like training for a test by memorizing the test. You might ace it and still be unable to change a tire. But even with skepticism, this is still a meaningful signal because the direction is consistent with everything we're seeing. The next leap is not only smarter models, it's better systems around models, better memory, better planning, better tool use, better loops, better supervision, better guardrails, better driver's seats. And that brings us to the practical takeaway. Don't wait for some mythical next model to change your life. Start thinking in workflows, start delegating outcomes, not sentences. Start building your own little harnesses with planning and iteration. Because the people who win in an agent world aren't the ones who can recite model names at parties. They're the ones who can get real work done faster, with fewer mistakes, and keep their hands on the wheel. That's what Nvidia's AVO result really points to. A future where the advantage goes to whoever can direct intelligence, not just access it. That's today's AI Inten. If you want to go deeper and learn AI with a community of people just like you, join us at aihammock.com. I'll see you tomorrow, my friends.