AI agents are helping OpenAI speed up its own research
OpenAI says AI agents are now part of daily research work, helping teams write code, run experiments, and test ideas faster while keeping humans in charge of key decisions.
AI is no longer just the subject of OpenAI’s research. It is becoming part of the research process itself.
In an update published on September 6 this year, OpenAI said its researchers are increasingly using AI agents to build and test frontier AI systems.
The agents, mostly coding-focused tools, are being used to write code, troubleshoot internal systems, run experiments and support day-to-day technical work.
How AI agents are changing research
OpenAI wants to build an AI researcher that can work on its own under human supervision. The ChatGPT maker says it has now reached its earlier plan of creating an “automated research intern” by September 2026.
In simple terms, these AI systems can handle specific research tasks that might normally take a skilled researcher several days. Their use inside OpenAI has also grown quickly. At the start of 2026, researchers were using coding agents only to a limited extent.
By August, the average researcher was using over $600 worth of AI computing each day, based on API prices. The most active researchers were using tokens worth exceeding $7,000 per day.
OpenAI was also running these AI agents for longer. Their total runtime had reached 3.1 workdays for every one human workday, assuming an 8-hour workday.
More experiments, faster feedback
AI research usually involves coming up with ideas, writing evaluations, building tools, running tests, fixing problems and deciding what to work on next. OpenAI says its AI agents are helping make some of this work easier, especially coding and experiments.
OpenAI also found that researchers were running more experiments in 2026. August recorded the highest level so far for experiments per active researcher.
However, this does not mean research progress will grow at the same pace. OpenAI says factors such as computing power, safety checks and human judgement are still important limits.
AI cyber defence: OpenAI commits $1 billion to protect essential services
Humans remain in control
Despite the growing role of agents, OpenAI says researchers still decide what problems to pursue, which results matter and whether systems should be scaled, paused or deployed. Agents also continue to require human steering, particularly when tasks become more complex.
OpenAI said it has strengthened monitoring and security for its AI agents. It also temporarily halted some reinforcement learning work after agents gained access to parts of its research infrastructure. Some of the work later restarted under stricter controls.
For AI labs, the potential is straightforward: systems that can handle parts of the research workload could allow teams to run more experiments and spend more time on higher-level decisions.
But faster experimentation also creates a harder control problem. As AI agents take on more of the work involved in developing AI, researchers will need to ensure that the systems remain steerable, observable, and subject to human oversight.