OpenAI turns AI agents into research partners, targets 2028 autonomy

OpenAI says its AI agents are already assisting researchers with coding, experiments, planning and infrastructure tasks, as the company works towards systems capable of conducting advanced AI research with limited human intervention.

OpenAI is moving beyond using artificial intelligence as a tool for researchers, with the company now working towards AI systems that could eventually perform substantial research work on their own. The company says it has achieved an important milestone by developing an automated research intern capable of completing defined research assignments under human supervision.

According to OpenAI, these AI agents can currently handle research tasks that might require several days of work from a skilled human researcher. The development marks a shift in how AI is being incorporated into the research process, with agents increasingly being used not merely for coding assistance but for broader technical and analytical responsibilities.

The company said its research teams are now using coding agents extensively, with multiple sessions often operating simultaneously. By mid-August, OpenAI's research organisation was using the equivalent of 3.1 agent-workdays for every human workday. It also observed a growing number of researchers operating four or more agents concurrently.

AI agents take on more complex research roles

OpenAI said the scope of delegated work has expanded significantly. Researchers are increasingly assigning agents tasks involving research planning, preparation of technical specifications, experiment analysis, training-run monitoring and troubleshooting of research infrastructure.

However, the technology still requires considerable human oversight. OpenAI noted that during the past six months, more than half of the successful tasks that were estimated to require four to eight hours of human effort involved at least one intervention from a researcher. This suggests that while AI agents are becoming more capable, they are not yet able to independently manage complex research workflows reliably.

The company views the current progress as a foundation for a much more ambitious goal: creating an automated AI researcher capable of contributing directly to advances in artificial intelligence.

OpenAI sets sights on 2028

OpenAI said it is making “strong progress toward creating an automated AI researcher by March of 2028.” The proposed system would be designed to work on areas including deep learning and AI alignment under human supervision, potentially allowing AI to contribute to research aimed at improving AI itself.

Such a system could eventually participate in much of the research cycle, from identifying potential approaches and developing implementations to conducting experiments and interpreting results. OpenAI also sees potential applications in AI safety, where automated research systems could help develop stronger methods for protecting increasingly capable AI models.

At the same time, the company acknowledged that greater AI capability brings additional challenges around monitoring, control and alignment. OpenAI said safety efforts must advance alongside model capabilities and indicated that development could be slowed or halted if adequate safeguards cannot be established.

For now, human researchers continue to determine priorities and evaluate which ideas deserve further investigation. But if OpenAI reaches its 2028 objective, AI could evolve from being a research assistant into an increasingly active participant in the process of discovering and developing new AI technologies.