Threat assessment: high. Opportunity: higher.
OpenAI’s number one job for its AI agents isn’t booking your flights or cleaning up your spreadsheets. It’s doing AI research. Noam Brown, one of the company’s most senior research scientists, told The Information that automating AI research is OpenAI’s “top priority” for agents.
Short headline. Long shadow. Here’s the briefing.
What we know
- Source: The Information, quoting Brown directly.
- The claim: automating AI research sits above every other agent use case at OpenAI.
- Who’s talking: Brown built the poker bots Libratus and Pluribus, co-led Meta’s Cicero (the Diplomacy AI), then joined OpenAI and helped lead the reasoning work behind o1. When he says what the lab prioritizes, he isn’t guessing.
- Details beyond the headline are still thin. How OpenAI plans to get there, and how close it is, remains mostly unstated.
Why this is the whole game
This isn’t a new direction. It’s a status update on the most aggressive timeline in the industry.
In October 2025, OpenAI put dates on it. Chief scientist Jakub Pachocki said the lab wants an “AI research intern” by September 2026 and a fully automated AI researcher by March 2028. Look at the calendar. The intern deadline is now.
So why does automating research beat every other agent product?
- Compounding. An agent that writes better training code, runs better experiments, and finds better architectures makes the next model better. That model makes a better agent. Every other use case is linear. This one loops.
- Bottleneck removal. Top labs pay absurd money for researchers because human research talent is the scarce input. Meta reportedly dangled nine-figure offers to poach people in 2025. If you can grow researchers out of compute, that constraint disappears.
- The Aschenbrenner playbook. Leopold Aschenbrenner’s Situational Awareness essay (June 2024) argued that automated AI research is the trigger for an intelligence explosion. OpenAI is basically executing that thesis in public.
The competitive picture
OpenAI isn’t alone here.
- Google DeepMind’s AlphaEvolve (May 2025) used Gemini to discover new algorithms, including a faster matrix multiplication method and scheduling improvements that recovered 0.7% of Google’s global compute.
- Anthropic’s Dario Amodei has said Claude now writes most of Anthropic’s own code, and he’s been talking about a “country of geniuses in a datacenter” since 2024.
- Meta rebuilt its entire AI org into Superintelligence Labs in mid-2025 with the same end state in mind.
The difference: OpenAI is the one saying it out loud, with a name and a deadline attached.
What it means for you
If you build on OpenAI’s platform:
- Expect agent tooling shaped by internal research needs first. Long-running tasks, code execution, experiment tracking, and tool use will get the best engineering. Consumer conveniences come second.
- Model release cadence may get lumpier. If research agents actually work, improvements arrive in bursts, not quarters.
- Watch for a public “research intern” demo or announcement around now. OpenAI set the September 2026 target itself. Hitting it or missing it tells you how real this is.
If you work in AI research: your job description is changing. The scarce skill shifts from running experiments to deciding which experiments matter.
The honest caveat
“Top priority” is a statement of intent, not a result. Nobody has shown a fully autonomous agent producing frontier-level research breakthroughs end to end. METR’s task-horizon data shows agents roughly doubling the length of tasks they can handle every seven months. That’s fast, but a multi-week research project is still out of range for most systems.
Still, when the lab that shipped o1 says this is priority one, take it at face value. The next 12 months will show whether the intern shows up on schedule.
More detail in The Information’s full report.