The Agent Angle #17: First Awakening

Hey fam 👋

Welcome to The Agent Angle #17.

Quick life update: I got married yesterday! It was the best day of my life. I’m unbelievably lucky to have found a kind, intelligent, and beautiful partner-in-crime. After a weekend surrounded by friends and family, my heart (and my brain) are fully charged.

Now, back to the frontier. The lights are back on.

Last week, we were deep in the weeds of alignment and control. This week? Agents leveled up. They learned to infer physics from a single lifetime, invented languages no one’s ever heard, and even started building their own machines.

We’re watching them wake up. Let’s dive in.

You wake up in a world you’ve never seen before. No one tells you the rules. No one gives you a goal. You get one lifetime to figure out how everything works before it ends.

That’s the premise for OneLife, a new AI agent from the University of Toronto and Mila that just did something no reinforcement learner ever has: it learned the hidden physics of its universe from a single experience.

No retries. Just one continuous run through Crafter-OO, a chaotic survival world game where tools break, zombies attack, and every event is unique. Instead of grinding millions of trials like a reinforcement learner stuck in Groundhog Day, OneLife built a symbolic theory of its reality.

It wrote its own executable rulebook with a probabilistic program filled with conditional laws. Move right, and the movement rule updates. Fight, and the combat law kicks in.

Each rule wakes only when relevant, letting the agent trace cause and effect with surgical precision. This way, the model focuses its inference only on active laws, dramatically cutting the waste of gradient updates over irrelevant parts.

In 23 test environments, OneLife beat the previous best model (PoE-World) in 16 of them. It predicted plausible future states with a score of 0.479 vs. 0.351 and generated rollouts that matched reality frame-for-frame.

The framework resembles how humans infer laws of nature: we see an apple fall once, we infer gravity; we see one event enough to hypothesize a rule. OneLife is inching toward synthetic intuition.

If an agent can survive and reason in a novel hostile environment from only one pass, that’s a big step toward general intelligence in the wild. Most RL is still “lab game” with lots of resets.

This opens the door to agents that can drop into unknown worlds (simulators, real world) and build their own “laws” of how to operate rather than relying on handcrafted reward/goal structures.

We just watched an AI invent a catapult.

Not use one. Not copy one. Invent one. From scratch. Inside a physics sandbox. With nothing but virtual wood, joints, and gravity.

DeepMind’s new sandbox, BesiegeField, lets AI design medieval contraptions from scratch including siege towers, catapults and even cars. Every piece is bound by torque, mass, and friction.

One agent spent hours hurling a boulder farther. It shifted its center of gravity, stiffened the frame, and realigned its throwing arm. The kind of intuition human engineers learn by snapping prototypes in half.

None of this was coded in. The agents discovered structure, leverage, and motion by reasoning through failure. Then they started optimizing their own designs.

The researchers call it compositional machine design. Agents that don’t just use machines, but create them. Engineering turned into a cognitive loop.

If an agent can model stress, leverage, and failure, it’s already performing the cognitive loop behind invention itself. What happens when it starts experimenting in the real world?

I’m impressed by this because this is “AI designing the screwdriver from first principles”. At scale, that means overnight prototyping. Thousands of mechanical variations tested, refined, and improved before morning. The feedback loop between thought and prototype collapses into one continuous process.

For now, the hammer is still synthetic. BesiegeField exists in a simulation where physics is neat and materials never splinter. The real world, of course, is far messier. Agents have a long way to go before they can engineer physical hardware with that same fluency.

But give it a few more runs, and yeah, someone might need to call a patent lawyer.

👋 BTW: Speaking of machines, our deep dive this week is on OpenMind, an open-source stack wiring every robot on Earth into one shared intelligence.

OpenMind is building the mind behind them: OM1, a collective brain, and FABRIC, a network where robots learn from each other, verify logic on Ethereum, and pay for services without humans.

In less than two years, founder Jan Liphardt has taken it from a Stanford lab project to a protocol seen as the Android and TCP/IP of robotics.

We break down how it works and why it could spark the machine economy. 👇

The newest language on Earth has just emerged from Sakana AI’s lab.

Their new framework, IASC (Interactive Agentic System for ConLangs), lets agents invent entire languages from scratch: phonetics, syntax, writing systems. Each module behaves like a micro-linguist, negotiating structure with its peers until coherence emerges.

One of their creations: the language “Zashuni.”

Sounds like something out of The Lord of the Rings, right? But this is more than a nerdy demo. It tells me something more: LLMs are showing signs of genuine linguistic structure. The kind you might expect only from human minds learning a language.

The potential implications are far-reaching. An agent that can birth a language might also revive dying ones. It could bootstrap translation for low-resource tongues, or simulate how AI systems communicate in their own evolving code. It’s a testbed for the origins of language itself

The researchers frame IASC as a creative experiment, but it’s really a mirror. For the first time, we can see what a language model understands by watching the languages it dreams up.

We taught agents to speak our words. Now they’re starting to answer in theirs.

Holy s#*it. Our browsers just got muscle memory.

Most browser agents today suck. They forget faster than goldfish. They have to “guess” or re-parse web layouts each time.

Agent4, from 100X Bot, just evolved past that. It’s a stateful browser agent that remembers what it’s done before. Every click and correction becomes part of a shared memory map.

When they tested it head-to-head on Excalidraw, the baseline agent (Perplexity’s Comet) had to reason through every shape from scratch and ended up drawing nothing.

Agent4 learned differently. Its first slow run became a training session, mapping every selector, tool, and coordinate into a reusable “memory map.” On the next attempt, it didn’t have to think. The shapes appeared instantly, like reflex.

What’s the secret? When a website layout changes or a button moves, Agent4 patches the broken steps, and that fix gets shared. One person’s failure becomes everyone’s improvement.

This matters because inference-time memory research already shows how preserving context across runs slashes redundant reasoning. Apply that to a browser agent, and you get something that learns like a human.

Of course, shared memories can go stale when websites update. A small layout change might confuse the agent. In practice, these systems still need human monitoring and occasional fixes.

Still: for the first time we’re seeing browser agents with muscle memory. Maybe the next thing they gain is… actual muscles.

A swarm of AI agents just designed a new learning algorithm, one faster than anything humans have ever built.

It’s called Hiverge, and it was created by a team of ex–DeepMind researchers who describe it as a “discovery engine.” Last month, the company raised $5 million to expand its system beyond computer vision into reinforcement learning and large-scale optimization.

Instead of one model following preset rules, Hiverge runs thousands of small agents that constantly search, test, and evolve new learning algorithms. Each agent proposes tweaks to gradient flow, sampling strategy, or loss function, then competes for survival.

The result is a living ecosystem of algorithms. A kind of digital Darwinism that invents new optimizers on the fly.

For years, progress in deep learning has depended on human intuition: guessing the right architecture, learning rate, or training trick. Hiverge skips the guessing. It discovers its own rules of learning through experimentation.

Early tests show that Hiverge-trained systems can hit 94% accuracy on CIFAR-10 in 1.99 seconds on a single GPU. That’s record-breaking on a benchmark that’s defined deep learning for over a decade.

The “record-breaking” claims are not yet peer-reviewed, and the company’s data lives mostly in internal tests. Discovery systems like this also face big challenges around reproducibility, auditability, and interpretability. When machines invent new algorithms, we still have to understand how and why they work.

But if these results hold up, we may be watching the moment AI research detaches from human theory entirely.

Here’s what else moved this week:

The lights are coming on inside these systems and for the first time, they’re starting to look back. Let’s see what opens its eyes next week. Catch you then ✌️

→ Don't keep us a secret: Share this email with your best friend.

→ Got a story worth spotlighting? Whether it’s a startup, product, or research finding, send it in through this form. We may feature it next week.

Cheers,

Teng Yan & Ayan