"Key Takeaways"
- OpenAI says coding agents are increasingly used across its research organisation, reporting 3.1 agent-workdays of effort for every human workday as of mid-August 2026.
- OpenAI also says researchers still set research priorities and judge results.
- Anthropic reports that as of August 2026, Claude "leads" 26% of its measured AI R&D work, with more than 90% at or above "AI collaborates."
- Anthropic explicitly says Claude was not fully autonomous for any measured subset of AI R&D work.
Summarize this article with
"AI building AI" sounds like a movie trailer. Inside the labs, it looks more like software work: coding, experiments, analysis, docs, and evaluations, done faster with agents sitting beside researchers.
Two of the biggest AI labs have now published internal measurements of that shift. The useful reading is not that researchers have been replaced. It is that AI research is changing shape, and the labs want the public to see the pace.
AI research is changing
Frontier labs used to treat models mainly as products. Now the same tools help build the next training run, the next eval suite, and the next safety check.
That feedback loop is why people talk about AI building AI. The phrase is shorthand for AI systems accelerating AI R&D, not for a model that sets its own scientific agenda alone.
What researchers are actually handing to agents
- Writing and editing research and engineering code.
- Running experiments and gathering results.
- Analysing logs, metrics, and evaluation outputs.
- Drafting documentation and internal notes.
- Helping with evaluations and routine engineering chores around model development.
OpenAI's research-agent numbers
In its research-acceleration write-up, OpenAI says coding agents are increasingly used throughout its research organisation. As of mid-August 2026, OpenAI reports 3.1 agent-workdays of effort for every human workday in that organisation.
Read that carefully. It is a measure of agent effort relative to human effort, as OpenAI defines it. It is not a claim that three robots replaced each researcher.
OpenAI also says researchers still set research priorities and judge results. In plain language: agents do more of the grind. Humans still decide what is worth grinding on, and whether the answer is good.
Anthropic's R&D automation numbers
Anthropic published a different yardstick: an R&D Automation Index built on Epoch AI's automation levels, from no AI involvement up to full autonomy.
As of August 2026, Anthropic reports that Claude "leads" 26% of its measured AI R&D work. In that framing, "leads" means the system can complete most of a task end to end from a high-level prompt while a human supervises.
Anthropic also reports that more than 90% of measured work was at or above the "AI collaborates" level, where AI can do large chunks under close human direction. And Anthropic is explicit on the ceiling that matters for public anxiety: Claude was not fully autonomous for any measured subset of AI R&D work.
What "AI building AI" really means
It means AI coding agents and research tools are inside the factory that builds models.
It does not mean a model woke up, hired itself, and shipped the next version with no human judgment. Both labs still describe humans as the people who aim the work and score the outcome.
Why oversight matters
Anthropic's same post spends as much energy on monitoring as on automation. It describes online and offline monitors over research and engineering agents, coverage rates, review latency, and escalation. The point is simple: if agents do more of the work, oversight has to keep up.
OpenAI's emphasis on humans setting priorities and judging results is the other half of the same idea. Speed without review is just a faster way to scale a mistake.
What this means for everyone else
If you are not training frontier models, you will not copy these exact dashboards. You will feel the side effects: coding agents that can stay on a job longer, research assistants that draft more of the first pass, and workplaces that expect one person to supervise several semi-autonomous threads.
The practical habit to borrow is measurement. Know which tasks you hand to AI, which ones still need a human signature, and how you notice when the agent wandered.
Frequently Asked Questions
Is AI already building AI on its own?
Not in the fully autonomous sense these labs are measuring. Anthropic says none of its measured AI R&D work was fully autonomous. OpenAI says humans still set priorities and judge results.
What does 3.1 agent-workdays per human workday mean?
It is OpenAI's mid-August 2026 report of agent effort relative to human effort inside its research organisation. It measures how much agent work sits beside each human workday, not a headcount replacement ratio.
Why should non-researchers care?
Because the same agent habits are moving into ordinary software and office work. The labs are simply publishing the numbers first.
The Chai Takeaway
AI research is speeding up because agents now do more of the middle work. Keep the drama down. Watch the measurements, keep humans on priorities and review, and borrow the oversight habit even if you never train a frontier model.
Explore More AI Updates