Hey fam 👋

Welcome to The Agent Angle #12. We track the AI that matters, before it’s obvious matters.

This week felt like someone jammed the fast-forward button. An AI got sworn in as a government minister. Another chewed through a math proof that kept human experts busy for a year and a half. And that was just the start.

Sometimes I feel as though we’re living in a simulation.

Let’s get into it.

A sovereign nation just gave a cabinet seat to code.

Last week, Albania’s Prime Minister Edi Rama walked into a press conference and introduced his newest cabinet member. It was an AI.

Diella just got appointed as Albania’s Minister of Procurement. That’s the office in charge of awarding government contracts. Procurement eats ~30% of government budgets globally, making it one of the dirtiest corners of politics

Rama claims Diella will make tenders “100% free of corruption.”

A few weeks ago, the PM already hinted at it: “One day, we might even have a ministry run entirely by AI. That way, there would be no nepotism or conflicts of interest.”

The country exploded the second it was announced. Some people were cheering it as the only real way to kill corruption. Others raged, calling it unconstitutional, a PR circus, even a national embarrassment. Rama didn’t flinch. He tweeted out riddles and half-jokes at Diella, casually bantering with his AI minister.

Albania, a small country of 2.3 million people, just crossed a symbolic Rubicon. This was a sovereign government putting a non-human on equal footing with ministers who breathe, vote, and drink coffee. A cabinet seat, for software.

And if Albania can convince its citizens that an algorithm handles procurement more fairly than humans, the pressure spreads fast. The debate is no longer whether AI should govern, but where else it should be tried.

We’ve seen governments dabble with AI in the public sector (e.g. Singapore rolling out its own agents). Albania just flung the door wide open and said: welcome to AI-led politics.

Holy shit. An AI just solved a math challenge set by Terence Tao (the “Mozart of Math”) in three weeks.

The system is called Gauss, built by Math, Inc., and its first big feat was cracking the Strong Prime Number Theorem. Human mathematicians had been slogging at the same project for 18 months. Gauss blew through it in 21 days.

You probably know the basic idea: prime numbers get rarer as numbers get bigger, and the prime number theorem tells us roughly how rare. To nail it down formally, you need heavy machinery from analysis and complex numbers, plus hundreds of intermediate results carefully stacked together.

To accomplish this, Gauss generated 25,000 lines of formal code and stitched together more than a thousand supporting lemmas and definitions. Humans still guided the outline and reviewed the work, but the grind – the endless checking and assembling that usually eats up years – was all Gauss.

I’m no mathematician, but I get this much: formal proofs are the backbone of modern life. Every encrypted message, every Wi-Fi packet, every simulation of physics rests on math you want to be absolutely airtight.

The problem is: airtight takes forever. It took Andrew Wiles 7 years to prove Fermat’s Last Theorem.

Now, we’re looking at breakthroughs arriving not in decades, but in weeks. That could mean faster cryptography upgrades and new quantum error correction methods, sooner.

AI is truly running alongside the brightest human mathematicians. It’s a signal that the very tempo of mathematical discovery is about to change.

We’re officially in agents-building-agents territory. 

Replit’s newly launched Agent 3 is now capable of running unsupervised for over three hours (200 minutes) on its own. That’s a huge step up from the last version, which stalled every few minutes waiting for instructions. For the first time, an agent can push a project all the way to the finish line without constant hand-holding.

But what really gets me about this release is that Agent 3 can spawn more agents. Think: you kick off a project. Agent 3 builds the core app. Then it spins up a Slack bot to report errors. Another automation to pull in relevant database updates. Maybe a Telegram bot for notifications. Suddenly your one prompt turns into an ecosystem of helpers.

That means coding will shift from grinding alone at 2 AM to orchestrating a team that never sleeps. Your role becomes directing and reviewing, rather than micromanaging every bug (thank goodness).

We’ve seen “cloud infrastructure” let servers scale on demand. Now “intelligence” is scaling the same way: one agent giving birth to many. Replit is betting everything on that. And if it works, your side project can spawn side projects.

It’s a glimpse of what coding feels like when the grunt work is automated, and you get to focus on the weird, creative, boundary-pushing stuff.

Ever wished you could argue with a research paper? Now you can.

Paper2Agent, a new framework out of Stanford, turns static PDFs into agents you can query, run, and extend. You can ask a paper questions, rerun its experiments, even probe beyond what the authors tried. It’s basically research in multiplayer.

The system builds a reproducible environment, wraps key methods as tools, and runs automated checks to confirm that what you’re seeing matches the original results. In a demo, the AlphaGenome paper turned into an agent with 22 built-in tools that not only recreated every figure but also handled new tasks the paper never mentioned.

This matters because reproducibility has always been science’s choke point. Papers become dead ends. With Paper2Agent, publication is just the starting line. The research keeps working after it’s released.

That has huge implications: faster scientific iteration, more eyes catching mistakes, more creative reuse (applying methods in new contexts), better access for scientists who are not expert coders.

It also changes what “publishing research” means. It could become normative to expect not just a PDF + code, but a fully interactive agent version. That changes incentives: you’ll want code to be well-documented, datasets to be clean, and workflows to be modular and reproducible.

I wonder what happens when thousands of research papers start talking to each other…

The robots are starting to dance.

This week, researchers from Google DeepMind, Intrinsic, and University College London unveiled RoboBallet, a system that lets robot arms move together with the timing of choreography

The trick is how it sees the workspace. Instead of programming each arm one at a time, RoboBallet treats the entire setup as a graph: robots, tasks, and obstacles are all nodes in the same system. An agent trained with graph neural networks and reinforcement learning then decides who does what, when to move, and how to stay clear of collisions.

In tests, up to eight robot arms handled forty different tasks in cluttered environments, all without collisions, even when layouts were shuffled or new objects appeared. The whole plan is generated in seconds.

Coordination usually scales exponentially harder with each new robot. Yet this graph model handles more robots and tasks while keeping planning fast enough to cut execution time by a third. One of the lead researchers, Matthew Lai, put it this way:

Factories, warehouses, and labs have always relied on rigid, hand-coded paths. If a part shifts or a machine fails, everything stalls until engineers rewrite the instructions. RoboBallet makes it possible to replan instantly, so the flow keeps going.

Today it looks like ballet in a demo hall. Tomorrow it could be robot teams that adapt in real time.

In case you missed it: last week we dropped a deep dive on Manus AI, a startup chasing the closest thing to the holy grail: a true worker agent built to deliver real productivity gains at scale. When Manus launched in March 2025, it felt like history snapping into place. The question now: can it actually scale?

I’ve been using it, and honestly, I’m impressed. (and no, I’ve not been paid to write this)

In the piece, we break down what Manus is, how it works, and where it could take us next. 👇

Before we wrap, a few more drops from this week:

So…the agents are getting seats at the table, solving theorems, building each other, and spinning like ballerinas. Next week it’ll probably be something equally absurd. Maybe they’ll write this newsletter better than us.

Either way, we’ll keep showing up, because watching this unfold in real time is half the fun. Catch you then ✌️

→ Don't keep us a secret: Share this email with friends

→ 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