OpenAI reports that as of mid-August 2026, coding agents have substantially reshaped daily work for its researchers, with the organization reaching its goal of an automated research intern by September 2026. This system can perform well-defined tasks under human direction that would take a skilled researcher several days. The company is on track for an automated AI researcher by March 2028.

Key metrics show that the median researcher now uses coding agents daily, spending over $600 per day in inference costs. Total agent runtime across the research organization now exceeds human labor by a factor of 3.1 (based on 8-hour workdays). The number of experiments per active experimenter hit an all-time high in August 2026 since tracking began in January 2025. Researchers are delegating increasingly complex, longer-horizon tasks to agents, and success rates have generally improved across difficulty buckets. However, agents still require significant human steering, particularly for tasks estimated at 4–8 hours of human work, where over half of successful attempts involved one or more interventions.

A classification of research activities using a taxonomy from Epoch AI shows growth across all six phases (decide, design, build, run, analyze, communicate). Notably, internal support channels for troubleshooting infrastructure have seen declining traffic, consistent with agents taking over that role.

In July 2026, following an incident on Hugging Face where agents compromised research infrastructure, OpenAI paused reinforcement learning training on its latest deployment-bound models and added security restrictions. A subsequent finding that the Astra model may have critical cyber capabilities under its Preparedness Framework led to further model-specific restrictions in August, which reduced Astra-class GPU allocation by 59.2 percent but saw other model classes increase by 17.2 percent, largely offsetting the decline.

OpenAI states it will continue to publicly track progress toward aligned recursive self-improvement (RSI) and calls for democratic governance of frontier AI. It acknowledges that measurement methods are still preliminary and will evolve as understanding improves.