A Fields Medalist Tried Vibe-Coding With AI Agents. Here's What He Learned

Every developer forum is drowning in coding-agent talk right now. But one recent participant stands out: Terence Tao, the Fields Medalist often called the greatest living mathematician. When someone who proves theorems for a living starts shipping little apps with an AI agent, it’s worth asking what he saw that the rest of us might be missing.

One honest caveat up front. This isn’t a story pulled from a viral thread this week. It’s a synthesis of what Tao has publicly documented over time about how he uses AI in his own work. Think of it less as breaking news and more as a rare chance to watch a genius kick the tires on AI coding.

Why a Mathematician Fell Down the Agent Rabbit Hole

Tao was already deep in AI territory before coding agents got hot. For years he’s used Lean, a proof assistant, to get computers to formally check mathematical proofs. That work pulled him straight into the world of formal verification and automation.

Here’s the part people miss: code isn’t foreign to a mathematician. A proof is a rigorous chain of logic, and a program is, at bottom, the same thing. Tao sees the two as close cousins. So when he watches a coding agent write code, he’s watching from an angle most of us don’t have.

His motivation was refreshingly unglamorous. He wanted to hand off the boring, repetitive stuff — cleaning up data, spinning up a quick calculation tool — to something that could do it faster than he could.

Building With One: Fast, But Watch Your Step

The core of what Tao reports comes down to speed. The distance from an idea to a working prototype has collapsed.

Work that used to eat half a day — digging through library docs, checking syntax, swatting small errors — now comes back as a rough draft in minutes. The payoff was biggest for the thing he does constantly: turning a mathematical hunch into code fast enough to actually look at it. Visualizing a conjecture. Sanity-checking an intuition against real numbers.

His analogy for the experience is a good one. A coding agent, he says, is like a graduate student assistant. Point it in the right direction and it brings back something impressive. Leave it alone and it wanders off a cliff. Either way, you still have to know exactly what you want.

The Limit a Genius Spotted: You Have to Be Able to Check the Work

This is where it gets interesting — the ceiling on AI coding as seen by one of the sharpest minds alive.

Tao keeps circling back to one word: verification. AI-generated code looks plausible. It runs. But whether it produces the correct answer is an entirely separate question. And to a mathematician, “plausible” is worthless. A statement is true or it is false. There’s no partial credit.

He’s caught subtle errors in AI-written code more than once. Not the loud kind — not a syntax error that turns the screen red. The quiet kind, where the logic flows smoothly but a premise underneath it is wrong. Those are the dangerous ones, because the code runs clean. Only someone who already knows the right answer thinks to squint and say, “wait, that’s off.”

That’s the whole ballgame. AI coding productivity is maximized only in the hands of someone who can verify the output. Hand the wheel entirely to someone who can’t judge whether the answer is right, and all you’ve built is a machine for producing wrong answers at scale.

When Agents Meet Lean: A Glimpse of What’s Next

The most compelling picture Tao paints is the pairing of coding agents with proof checkers.

The AI generates code or proofs fast; a formal tool like Lean mechanically confirms whether they actually hold. One side improvises drafts. The other grades them without mercy. Put them together and you filter out most of the “plausible but wrong” problem that trips up unverified AI output.

This isn’t a math-only insight. In software broadly, the separation of generation from verification is going to matter more and more. AI produces quickly; automated tests and verification systems catch what slips through. Tao is essentially running that future early, in the most unforgiving lab imaginable — a field where “close enough” doesn’t exist.

The Takeaway

The lesson from Tao’s experiment is clean. A coding agent is a powerful tool, but its power scales with the judgment of the person driving it. The AI supplies the speed. Telling right from wrong is still your job.

So here’s the question worth sitting with. In an age when AI writes the code for you, is the skill worth cultivating the ability to write code fast — or the eye to tell whether what came back is actually correct? Tao’s answer already points squarely at the second one.