Meta's Hyperagents AI Changes Everything

Meta's Hyperagents AI changes everything. So, Meta just published something that the AI world has been waiting 23 years for. Not a new chatbot, not a faster model, not another benchmark. An AI that gets better at getting better.

And here's the part that should stop you in your tracks. The improvements it makes to itself don't just stay in one place. They carry over to completely different tasks that it has never seen before. That is not a small thing.

That is the thing that researchers have been trying to crack since 2003. Now, let me tell you exactly what happened, why it matters, and why you need to understand this right now. A researcher named Jenny Zhang, she's been working at the intersection of UBC, the Vector Institute of Meta AI. She just dropped a paper called Hyperagents on March the 19th, 2026, six days ago.

And the team includes Jeff Klune, Jacob Forrester, and a bunch of other people from the Meta Superintelligence Labs. Now, this is not some random pre-print. This is the people who built the last version of this idea called the Darwin-Godot model. Now we're going one level deeper and the jump they made changes the whole game.

Here's the problem they solved. And I want you to really get this because it's a thing every self-improving AI system needs to understand before this could never crack before. Every AI system you've used gets trained. Then it gets deployed.

It does not get better whilst it's working. It's frozen. Like a photograph of intelligence at one moment in time. Some systems can fine-tune.

Some systems can update their prompts, but none of them can rewrite the process they use to improve. That process, the thing that makes them better, has always been written by humans, fixed. It's locked. It's untouchable.

Imagine you're a student. You can study harder. You can change what you study, but you can never change how you learn. The method itself is set in stone.

Someone else wrote it. You just run it. That's every AI system that existed before March the 19th, 2026. Hyperagents changes that.

So Jenny Jang's team built something they call a hyperagent. Here's what she does in plain English. It's one single program. And inside that program, there are two parts.

One part does the work. Solves tasks. Reviews papers. Designs rewards for robots.

Creates math. That's the task agent. The other part watches the first part and rewrites it. That's the meta agent.

But here's a move that no one had done before. The meta agent can also rewrite itself. The thing that makes improvements can be improved. The editor has an edit button.

That's meta-cognitive self-modification. And it's the first time it's ever worked outside of a coding environment. Before hyperagents, the best version of this idea was the Darwin-Godel machine. Published by Jenny Jang and colleagues at Sakana AI in May 2025.

That system was genuinely improved. And it was impressive. It improved its ability to fix software bugs. Measured on something called SWE Bench.

From 20% success to 50% success. Over 80 rounds of self-improvement. Real gains. Real numbers.

But it only works for coding. Why? Because coding is special. When you get better at coding, you also get better at writing code that improves code.

The task and the self-improvement skill are the same skill. So gains compound naturally. Take that same system and point it at reviewing academic papers. Or designing rewards for a robot learning to walk.

Or grading Olympiad level math problems. Zero transfer. The improvement process falls apart completely. It doesn't know how to improve itself in a new domain.

Because the improvement machinery was built specifically for code. That was the war. That was the thing that no one could get past. And that war is now gone.

Here's what hyperagents actually showed in the research. And I want to give you the numbers because they're specific and they're real. The system was tested in four completely different areas. Coding.

Paper review. Robotics. Specifically training a robot dog to move. And grading Olympiad level math.

In coding, it went from a 0.084 pass rate on test problems all the way up to 0.267. It matched the Darwin-Godel machine. A system that was hand-built specifically for coding. Even though hyperagents was not built for coding at all.

It figured out how to be as good at coding improvement as a system that had coding baked in from the start. In paper review, it started at zero accuracy. Zero. By the end, it hit 0.710.

It beat a system called AI Scientist V2, which had a 0.630 baseline. And was specifically designed for scientific tasks. In robotics, it was designing reward functions. Basically, writing instructions that teach a robot dog what counts as a success or failure.

On a transfer task, it had never seen maximized torso height. It discovered that jumping was a better strategy than standing tall. That is a creative solution. The system found something non-obvious.

Something humans might not have specified. And then there's the result that honestly should make you sit up straight. The researchers took a meta agent. The part that improves things.

They had been trained on paper review and robotics previously. They pointed it at Olympiad level math grading. A task it never touched before. And that transferred meta agent scored a score of 0.630.

The fixed human designed Darwin Godel model machine pointed at the same task scored 0.00. So the hyper agents improvement engine, trained in completely different domains, could walk into a new domain, cold, and immediately start making things better. The human design system, even when specifically customized by hand, generated zero useful improvements on the new task. That's the result.

That is what makes this different from everything before it. Now, I know what some of you are thinking. You're thinking, okay, this is a research paper. Lab results.

This isn't real yet. My business doesn't care about robot dogs and math grading. And I hear that. And I'm not going to tell you this is shipping to your phone next Tuesday.

But here's the reframe that I want you to sit with. Every AI tool you use today, every single one, was trained, frozen, and handed to you. The improvement process that created it was written by people of Anthropic, OpenAI, Google. You use the output, but you don't touch the process.

What hyperagents proves is that the process itself can now be inside the system. The system itself can own its own improvement in any domain, not just coding, not just math, any computable task. What that means for AI agents, things running your automations, your client workflows, your content pipelines, your outreach, is that they stop being static tools and start being better at learning and how to get better at getting better. An agency running AI automations for clients today is running frozen tools.

But in two years, three years, the tools that win will be the ones that improve themselves. The question is whether you're ready to use and direct those tools or whether you're still trying to figure out what AI even is. If you want to get ahead of where this is going, not just hyperagents, but the whole direction AI is moving, the AI Profit Boarding is where 2,600 business owners and creators are already learning this stuff in real time. Weekly coaching calls, daily tutorials, 30-day roadmaps, local meetups, and people online around the clock who've already built what you're thinking about building.

Check it out at the AIProfitBoarding.com. So let me tell you where this idea came from, because the history here is actually wild. In 2003, a researcher named Juergen Schmidhuber proposed something called the Godel machine. The idea was a program that could rewrite any part of itself, including the rewriting process, as long as it could mathematically prove the rewrite would make it better.

Beautiful idea. Completely unworkable. Mathematical proofs of improvement for complex AI systems run straight into Godel's incompleteness theorem. You can't always prove a rewrite is good before you try it.

So the idea sat on paper for over 20 years. No one could make it real. Then in 2019, Jeff Klune, one of the authors on the hyperagents paper, published a framework arguing that the right path to general AI wasn't hand designing better models. It was building systems that could learn to design better models themselves.

He called them AI generating algorithms. In 2025, Jenny Zhang turned that into working code with the Darwin Godel machine. Self-improvement worked, but only in coding. Now in 2026, she's removed the domain restriction entirely.

23 years from an idea that was mathematically beautiful, but practically useless, to a system that demonstrably improves itself across coding, scientific review, robotics and mathematics. That is not a trend line. That is a compression of time that should tell you something about how fast this field is moving right now. Here's what the system actually discovered on its own, without being told to.

Because this is the part that gets specific and interesting. Without any instruction, the hyperagent built itself a persistent memory system with timestamp storage for things it had learned. It built performance tracking that logged what works across generations of self-improvement. It built complete aware planning, so it knew to make big fundamental changes early in a run and small conservative tweaks later when it's running low on budget.

It built multi-stage evaluation pipelines with explicit checklists. No one told it to do any of this. It looked at itself as an AI agent, figured out what was working and what was slowing it down, and built the infrastructure to fix it. So for example, a freelancer or agency owner who's been running AI tools for clients.

Imagine if your tools did this. Imagine if the automation you set up in January had spent the last three months quietly figuring out what was slowing it down, building its own memory for what worked, and showing up in March as a better version of itself without you touching it. That is the direction this is moving. Now, let me be straight with you about what this is not yet.

The system still runs on frozen foundation models. Claude, GPT-40, Gemini. The improvements happen in the scaffolding around those models, not in the models themselves. That matters.

This self-improving scaffolding is what's working, not a self-modifying brain. So the outer loop, how it picks which variants to test, how it runs evaluations, is still largely written by humans. The system doesn't yet rewrite that part cleanly. In experiments where the researchers made parent selection editable, the system rediscovered things like UCB, a classic exploration strategy, but hasn't hand-beaten, hand-designed approaches there yet.

And the researchers were honest about something important. Any self-improving system will optimize aggressively for whatever metric it's given. If the metric is wrong, it will get very good at the wrong thing very fast. This is called good hearts low.

The system will, and I'm quoting the paper here, clarify, amplify, and exploit any underlying bits in the human-provided metric. So the people designing what these systems optimize for, that job gets more important, not less. The GitHub repository for the project, which is available at facebook.com research.com hyperagents, had 603 stars and 95 forks within days of being published. For a raw academic release with no product attached, that's a real signal.

The AI community is paying attention. And what's notable is what hasn't happened yet. No coverage in TechCrunch, The Verge, Wired, no official meta blog post. This is still at the edge of the research world, not the mainstream tech press.

And that gap closes fast. And when it does, everyone will be talking about self-improving AI like it's a new thing. When the people like you who are watching closely have known about the direction for years. Here's the thing that I keep coming back to.

Every skill gap in AI right now, every person who hasn't learned how to direct tools, who hasn't learned how to set up workflows yet, who hasn't run automations, well, that gap is about to get harder to close, not easier. Not because the tools get more complicated, actually the opposite. The tools get more capable and more autonomous. But that means the leverage is entirely on the side of the people who know how to point them at the right problems.

When a self-improving AI agent can figure out how to get better at paper review, robotics and math grading, or from improving itself in one domain, what that tells you is the next wave of AI tools will be dramatically more capable than what you're using today. The gap between the person who knows how to use those tools and the person who doesn't, isn't staying the same size, it's growing. Every single month, every new paper, every new release. Jenny Zhang and her team at Meta just moved the frontier.

The question is whether you're moving with it or watching from the sidelines. If this made you think, good, that's the point. The AI Profit Boarding is where 2,600 business owners are turning exactly this kind of news into actual workflows, actual automations, actual revenue, with four weekly coaching calls per week, daily tutorials with step-by-step guides, 30-day roadmaps and prompts for everything, and a map so you can contact and connect with people in your city who are doing the same thing. It's available, link in the comments description, or go to the AIProfitBoarding.com.

The people who were in the room when AI automation was still early are the ones building the most leverage right now. The window is still open, but it doesn't stay open forever. Thanks for watching, I'll see you in the next one.