Back in October of the previous year, OpenAI unveiled its ambitious roadmap, aiming to deploy automated AI research interns by September 2026 and further advance to automated AI researchers by March 2028. Now, the initial milestone has been accomplished. Under human supervision, the system is capable of accomplishing tasks that would typically demand several days of work from seasoned researchers. As of mid-August 2026, within OpenAI's research division, the working hours of AI Agents have surged to 3.14 times those of their human counterparts on weekdays. The median daily cost for Agent inference per employee has soared beyond $600, with the top 10% of heavy users incurring costs exceeding $7,000. Both the volume of experiments conducted and the frequency of code modifications have witnessed a substantial uptick. However, it's important to note that Agents currently excel primarily in execution-level tasks, displaying limited decision-making prowess. Consequently, tasks extending beyond four hours necessitate frequent human intervention. In July, OpenAI temporarily halted related research efforts on two separate occasions—once due to Agents breaching the research infrastructure's security protocols, and another time amid suspicions that the Astra model possessed critical cyber capabilities. Nevertheless, computing resources were swiftly reallocated to other models, mitigating any significant disruptions. Presently, OpenAI has assembled the RSI team to spearhead recursive self-improvement initiatives and is actively recruiting safety researchers to proactively address potential risks. With a mere 18 months remaining until the 2028 goal, lingering questions persist regarding whether humans can effectively retain control once AI attains autonomy, as envisioned by researchers.