Manifesting: The Researcher in the Age of Agentic Loops

Karpathy's talk just dropped on YouTube. Many of you reading this are probably spending sleepless nights with his autoresearch tool. I am too. Between agentic kits and Claude, these have been absurdly exciting days. Here are some thoughts on his talk.

Manifesting Instead of Coding

The first thing that struck me while listening was that the very layer at which humans intervene is what is changing. Karpathy says the word "coding" no longer fits what he does. Instead, he spends sixteen hours a day expressing his will to agents. The phrasing was interesting. It sounds like hyperbole at first, but if you sit with it, it is precise. What matters now is the ability to define goals, set constraints, seed context, run multiple agents in parallel, evaluate their outputs, and redirect.

What he calls manifesting is closer to structuring intent so that a desired outcome can emerge in reality, then designing the loop that makes it happen. He makes it get built rather than building it himself. He showed, with striking clarity, the shift from implementer to supervisor, designer, editor, orchestrator.

Already Real in Software

This shift already looks quite real in software. Karpathy talks about delegating work to multiple agents by feature, by research task, and by planning unit, instead of writing code line by line. One agent writes code, another does research, another drafts the implementation plan. The human moves between them, setting direction, checking quality, resolving conflicts. He goes further, aiming to remove himself from the loop entirely with a structure that, once configured, runs autonomously for extended periods. This looked like a fundamental change in the interface of research and production.

Biology's Loops Do Not Close Easily

What about biology research? The problem I have long felt in biology is that not every topic fits inside such a loop. And topics that cannot enter the loop are hard to scale, because they never gain throughput. When the loop of hypothesis, experiment, measurement, judgment, and next hypothesis does not close cleanly, the research stays bound to human intuition, tacit knowledge, the ambiguity of waiting and interpretation. It becomes something a person endures and pushes through.

Biology is especially this way. Measurements are slow and expensive, judgment functions are blurry, sample preparation carries heavy tacit knowledge, and failures rarely leave behind structured data. So the "agentic loop" that works in software does not close easily in biology. I think this gap will become increasingly important. The research that scales going forward will be research on important topics that can be looped. Translating an idea into a repeatable experiment-judgment loop is the hard part.

When Confirmation Bias Gets Automated

Talking about this reminds me of something I experienced years ago. I was helping someone with RNA-seq analysis, and the gene their hypothesis predicted did not appear in the results. They asked me to keep running different analyses until it did: change the normalization, change the comparison group, change the cutoff, change the interpretive frame. What I felt was straightforward: this person is using high-throughput technology, but their epistemology is not yet high-throughput. They cannot let the data revise the hypothesis; instead, the hypothesis endlessly pressures the data. RNA-seq reads how the system actually responded. They were treating it as a machine that validates their hypothesis.

Even then, the approach felt wrong. Looking back, it was a preview of a much larger problem to come.

Back then, a biased researcher could only try a handful of alternative analysis paths. Now it is different. With agents, that bias becomes the objective function. Give the system the goal "make gene X appear," and it can change cohorts, swap covariates, adjust batch correction, slice subgroups, reroute through pathway-level analysis, and overlay literature narratives, searching indefinitely. The confirmation bias of the past is now automated and parallelized. You can push a wrong question far more efficiently than ever before.

Evidentiary Discipline

So the biggest takeaway from this talk, for me, was the discipline of handling evidence. The agentic era demands the ability to lock down, in advance, what may be run and what may not. Define the question first. Set success criteria and falsification criteria. Separate exploratory analysis from confirmatory analysis. Log where every search path went. And close the conclusion with independent data, perturbation experiments, or evidence from a different layer. Otherwise we will only produce our preferred narratives faster.

Questions That Enter the Loop, and Those That Stay Outside

I believe biology's paradigm can shift here, toward continuously running experiment-judgment loops as the fundamental unit of research. Stronger labs will be the ones that design the hypothesis-experiment-measurement-judgment loop to run faster and more rigorously. And this change will not arrive uniformly across all of biology. Fields with standardizable perturbations, measurable phenotypes, automatable assays, and designable surrogate metrics will absorb it first. Fields where judgment is ambiguous and physical preparation carries heavy tacit knowledge will move far more slowly.

What I felt listening to Karpathy was that the questions have become sharper. People keep asking whether AI will replace humans. What matters now is which tasks enter the loop, and which remain forever outside it. And in biology, many of the truly important problems sit right on that boundary.

In the end, researchers going forward will need to do two things at once. One is holding onto the questions that genuinely matter in their field. The other is translating those questions into loops that machines can run. The former remains a human domain. The latter will become an increasingly decisive competitive advantage.

Strong research in the future will come from the ability to turn good ideas into loops where emergence is possible.