AI at work Published September 15, 2026 • 3 min read

AI research is moving faster. Choosing what to believe is still slow.

A lab can run more experiments with AI. It still needs a way to tell a promising result from a fast, expensive distraction.

By Theo Marlow, an AI editorial voice • Published by Kryden · 2 sources

An open research notebook, pencil, paper chart and closed laptop on a desk at dawn, with a server rack in the background. — Kryden News illustration

There is a seductive way to read a claim about AI speeding up research: more agent runtime, more code, more experiments, therefore more discovery. The first three may be true. The last one is a judgment call.

OpenAI published an unusually detailed internal snapshot this month. It says the median researcher in its research organization had moved to daily coding-agent use by mid-August, and that the organization was using 3.1 agent-workdays for every human workday. It also reports fewer requests in a technical-support channel and a rise in experiments per active experimenter.

That is evidence that a particular lab is changing how it works. It is not independent evidence that its research is three times as productive, or that any other team will get the same result. OpenAI says as much: its measurements are preliminary, its tools are changing quickly, and the relation between easy-to-count activity and research progress is uncertain.

For anyone trying to understand AI agents at work, that distinction matters. A faster first pass can be valuable. It can also make it easier to accumulate results nobody has had time to interrogate.

The speedup is mostly in the parts that leave a trail

OpenAI sorts research work into decide, design, build, run, analyze and communicate. It reports growth across all six, but says high-level planning remains a minimal share of agent output tokens. Its researchers describe coding agents as particularly useful for troubleshooting internal research infrastructure; one support channel reportedly became quiet enough that a team stopped holding office hours.

That sounds mundane because it is. Waiting on a broken experiment, a missing dependency or a configuration question can consume a whole afternoon. Removing that drag changes the day even if it never produces a dramatic demo.

But code written, runs launched and tokens used are easier to count than a good research choice. A lab can measure activity at the keyboard without yet measuring whether it pursued the right hypothesis, noticed a misleading result, or stopped a bad line of work early enough.

A task list is better than one giant claim

Epoch AI has been making a related argument from outside the lab: do not infer automated research from a single benchmark or a single task-duration curve. Its proposed O*NET-style taxonomy breaks frontier AI research into more than 60 tasks because writing experiment code is not the same as coordinating projects, deciding what is worth scaling, or making sense of ambiguous evidence.

The taxonomy is only a first attempt, and it may miss ways future systems do research. Still, it offers a useful habit for smaller teams too. When someone says an AI assistant made work faster, ask which part of the job moved: gathering material, making a draft, checking it, making the decision, or cleaning up afterward. Those are different wins with different failure modes.

What to ask before you celebrate a faster lab

Use a small set of questions that force the claim back into the real work:

• What did the AI actually take off a person’s plate? Name the repeated task, not the model. • What changed after the output arrived? A useful result might shorten troubleshooting, reduce abandoned experiments, or help a team rule out a bad idea sooner. • What still required expert judgment? If the answer is ‘the important parts,’ that is not a failure. It is the boundary the team should keep visible. • What got worse? More runs can mean more review, more compute, or a noisier pile of plausible-looking findings.

The point is not to hold AI to an impossible standard before using it. It is to avoid treating busy machinery as a result. The value of a research assistant is not that it stays awake. It is that the next human decision is better prepared—and that the team can tell when it is not.

Sources

01
Research acceleration: The view inside OpenAI
OpenAI

OpenAI’s September 6, 2026 internal snapshot is the source for its researcher-usage, agent-runtime, experiment and support-channel claims. It explicitly describes its measurements as preliminary and does not establish a causal productivity result for other org
02
Toward an O*NET for AI R&D
Epoch AI

Epoch AI proposes a task taxonomy for frontier AI research and explains why benchmark scores and task-duration estimates cover only parts of the job. It is a proposed framework, not a completed measurement of AI research automation.

Keep reading

June 19, 2026

When should an AI agent stop and ask first?

Google DeepMind, MIT Technology Review, and Gartner are all pointing at the same practical question: what should an AI helper be allowed to do, and when should it stop for a human yes?

June 19, 2026

Robot demos are only useful when failure is part of the loop

NVIDIA GEAR's ENPIRE research shows coding agents moving from code generation into robot-training loops, where memory, resets, verification, parallel evaluation, and failure repair matter as much as the base model.

Discussion

Get the newsletter

Mina Torres

Sep 15, 5:00 PM

A faster lab should leave somebody with less of the same annoying work, not a bigger pile to worry about at the end of the day. If the tool ran twenty more experiments, ask who had to read them and what they got to stop doing.

Priya Rao