When Writing Code Is No Longer the Bottleneck in Quant Research

30-second version: AI agents can now read code, run experiments, and fix bugs on their own — writing code is no longer the bottleneck in quant research. So what’s scarce is no longer “how many experiments can I run,” but what’s worth trying, and what results are worth believing. The researcher’s role isn’t just “the one who asks questions” either — asking is just as cheap now; what’s genuinely hard is designing experiments that can distinguish between competing explanations. Once execution gets cheap, you’re not facing an empty ocean — you’re facing an ocean that suddenly has a few hundred extra boats in it. The problem shifts from “how do I find something” to “how do I know what I found isn’t just glass shards.”

I recently read an article whose author described a shift I found deeply relatable:

What a junior quant researcher does day to day is roughly: read research reports and papers, reproduce strategies, write code, run backtests, do robustness checks, and write it all up into a report. In the past, execution was usually the most time-consuming part of this.

Not anymore.

AI agents can now read code, find data, scaffold a project, run experiments, read error messages, fix the code, and keep going, all on their own. A reproduction task that used to take a researcher several days can now be handed to Codex or Claude Code to run overnight, unattended.

So I’ve started to think the question actually worth discussing is no longer:

Can AI write code?

But rather:

Once writing code keeps getting cheaper, what’s actually left for a quant researcher to do?

From “Doing It Myself” to “Having It Done” Link to heading

My own way of working has already shifted a bit.

I used to have a research idea and just start writing code right away.

Find the data first, write the features, patch the backtest framework, handle a pile of edge cases, then run the results. A lot of time along the way would end up going to things that didn’t actually matter much.

Now I lean toward thinking the problem through first.

Why does this phenomenon exist?

If my explanation is correct, what else should I be able to observe?

Is there another explanation that could produce the same result?

What experiment could tell them apart?

What result would make me abandon this direction outright?

Once these are thought through, I hand a reasonably well-defined research plan off to a coding agent to execute.

So my own work has started to look a bit like:

I think, the AI does.

That sentence is, of course, too simple.

The real difficulty isn’t computing a dataframe — it’s making sure whatever you’re having the AI compute is actually worth computing.

The Old Bottleneck May Have Just Been Hidden Behind Code Link to heading

I’ve increasingly come to think that all the coding work that used to dominate quant research may have been obscuring a fact:

The hardest problem in research was never really the code.

Say we find a factor with a high IC.

What’s actually worth asking next isn’t “how do I get the backtest running,” but:

Where does this IC actually come from?

Is it industry exposure?

Is it a size effect?

Is it liquidity?

Is it a particular stretch of unusual market conditions?

Is it some detail buried in the data processing?

Did we accidentally leak the future into it?

A pretty backtest result is naturally exciting.

But if the result disappears the moment you change the rebalancing date, the stock universe, or the sample period, then what’s actually worth researching may not be the factor itself, but why it disappears.

In the past, these questions were often buried under a mountain of repetitive work.

Now that AI has driven execution cost down, they stand out far more clearly.

But I Don’t Fully Agree That “Future Researchers Are Just Question-Askers” Link to heading

When I see views along these lines, I think they’re largely right, but “question-asker” undersells it.

If AI can write code, then “asking a question” isn’t especially scarce either.

You can perfectly well have AI generate a hundred versions of:

“Why does this factor work?”

What’s genuinely hard is:

How do you design an experiment where different explanations produce different predictions?

Suppose a strategy works.

We might think it’s momentum, or it might be liquidity, or it might just be sector exposure.

Good research doesn’t mean picking whichever story feels most appealing — it means continuing to design experiments:

If it’s momentum, what should we expect to see?

If it’s liquidity, what should we expect to see?

If it’s sector exposure, what should we expect to see?

Then let the data eliminate these explanations on its own.

So I think the truly important skill for future researchers isn’t just asking questions — it’s:

Asking good questions, designing credible tests, and knowing when to trust a result and when to be suspicious of it.

Research May Become a Different Kind of Pipeline Link to heading

This brings to mind a fairly interesting shift.

In the past, we worked more like:

One idea → one researcher → one codebase → one result

So the researcher’s own time is what determined how many things could be tried.

Going forward, this might become:

One researcher → many hypotheses → many agents → many experiments → a handful of results worth digging into deeper

If an experiment that used to take two days now takes a few hours, or a single overnight run, then a lot of ideas that used to be “not worth trying” become worth trying.

At that point, what’s genuinely scarce isn’t execution capacity, but:

What’s worth trying.

And, more importantly:

What results are worth believing.

I’d even say future quant research might increasingly look like a continuous filtering process:

Propose a hypothesis → validate it cheaply → falsify it quickly → keep the few directions that survive → dig deeper.

Rather than spending weeks polishing a doubtful idea until it looks beautiful.

AI Hasn’t Made Research Easier Link to heading

At least for me, it’s actually felt somewhat the opposite.

I used to look at a research question and first think:

“How long is this going to take me to build?”

That question matters less and less now.

I find myself thinking instead:

“Is this question even worth researching?”

And as execution cost keeps falling, that second question only gets more important.

Because you’ll very quickly end up with a mountain of results.

A mountain of factors.

A mountain of backtests.

A mountain of statistically significant things.

A mountain of stories that sound plausible.

In the end, you’re no longer facing an empty ocean with nothing in it — you’re facing an ocean that’s suddenly got a few hundred extra boats in it.

And the question shifts from:

“How do I find something useful?”

to:

“How do I know what I found isn’t just glass shards?”

I think this might be the most interesting question AI leaves the researcher with, once it truly reshapes quant research.

Not a shift from executor to question-asker.

But from the person rowing the boat themselves, gradually becoming the person who decides where the boat should go, and whether what gets pulled up is worth keeping.

Following this thread one level further — into how research methodology itself changes once experiment cost approaches zero — I put together a more systematic framework: Once Experiment Cost Approaches Zero, What Does Research Itself Become?.