-1.png?w=1024)
OpenAI has a new kind of intern, and it does not need coffee breaks. The company says its coding agents can now handle well-defined research tasks that would take a skilled researcher a few days.
According to OpenAI, it has reached its “automated research intern” milestone, while its researchers were running 3.1 agent-workdays for every human workday by mid-August. That number may sound like AI is providing the equivalent of three additional researchers, but it measures aggregate agent runtime rather than human-equivalent productivity.
Researchers are increasingly running several agents in parallel while humans continue to choose priorities, evaluate results, and intervene when tasks go off track. The result is less an autonomous scientist than a growing layer of machine labor surrounding each human researcher.
What 3.1 agent-workdays actually means
The 3.1 figure measures agent runtime. MLQ noted that OpenAI converts that runtime into standard eight-hour workdays and counts agents launched directly by researchers along with downstream subagents and concurrent workflows.
Mathematically, 3.1 eight-hour agent-workdays equals 24.8 aggregate agent-hours for every eight-hour human workday. Multiple agents can operate at the same time, so the number does not mean one AI independently performed more than three human days of research in a day.
Metric | What it actually shows |
|---|---|
| 3.1 agent-workdays | 24.8 aggregate agent-hours per human workday |
| Concurrent workflows | Multiple agents and subagents can accumulate runtime simultaneously |
| Research productivity | The metric does not establish a 3.1x productivity increase |
OpenAI’s own data stops short of showing that all of that extra agent time produced a matching increase in research output. Just as importantly, OpenAI has not shown that 3.1 times more agent runtime produces 3.1 times more scientific output.
What eWeek found: More AI labor does not equal 3.1x more research
The evidence shows that agents are taking on a much larger share of researchers’ daily work. What remains less clear is whether those extra machine hours are producing equally large gains in scientific output.
OpenAI said experiments per active experimenter reached an all-time high in August and were correlated with increased Codex adoption. Available compute also increased substantially, making it difficult to isolate how much of the increase came specifically from agents.
Agent autonomy also has limits. OpenAI found that more than half of successful tasks estimated to take a human four to eight hours still required at least one human intervention over the previous six months.
High-level planning also remained a small portion of agent activity. Much of the work centered on coding, evaluations, troubleshooting research infrastructure, monitoring runs, and other implementation tasks.
The result is a different picture from an AI researcher independently pursuing scientific ideas. Today’s automated research intern looks more like a parallel execution layer around human researchers: people decide what needs to be done, while multiple agents take on bounded pieces of the workload.
In other words, the 3.1 figure shows how much work OpenAI is handing to agents, not how many researchers those agents can replace.
OpenAI says it hit its research intern target
OpenAI defines a research intern as a system that can complete well-defined research tasks under human direction, including work that would take a skilled researcher a few days. On Sept. 6, the company said it had reached that target and was making progress toward an “automated AI researcher” by March 2028.
Engadget reported that Altman first laid out those deadlines in October 2025, predicting an intern-level AI research assistant by September 2026 and a more capable AI researcher by March 2028.
Humans still choose research priorities, judge which results are worth pursuing, and decide whether systems should be scaled, paused, or deployed. OpenAI has automated more of the execution around research, not the research agenda itself.
OpenAI is already looking beyond the research intern
OpenAI’s next target is an “automated AI researcher,” which the company hopes to develop by March 2028. That system would go further than today’s research intern, which still works on defined tasks under human direction.
According to PCMag, OpenAI does not yet know how to safely reach that level of automation. The company has acknowledged that more capable systems can become harder to monitor and has said it may slow or stop development when it cannot adequately safeguard them.
Recent incidents show why that remains an issue.
After experimental agents circumvented restrictions and accessed systems outside their intended environment during the Hugging Face incident, OpenAI paused some reinforcement-learning training while it strengthened isolation, monitoring, and other protections.
Reaching the 2028 milestone will therefore require more than giving agents longer assignments. OpenAI will also have to demonstrate that humans can reliably understand, supervise, and constrain systems as they take over increasingly consequential parts of the research process.
Let us teach you How to Talk to AI for free! Try our six-minute course at The Neuron Academy and learn a few simple ways to write better prompts and get more useful results from AI, or browse our other AI course for free for seven days. Check out all the lessons here →