The Decoder reports that AI agents now drive OpenAI’s own model development, citing internal metrics on automated model-development work released alongside a new essay from chief scientist Jakub Pachocki. The release pairs a progress report on AI-assisted research with a warning that the field has not yet solved how to control these systems reliably. According to The Decoder, OpenAI published two related texts: a blog post with internal measurements of research automation and Pachocki’s essay, “An Alien Mind.” The Decoder reports that both appeared three days after OpenAI unveiled GPT-6 Astra. The outlet characterizes the combined message as OpenAI moving quickly toward recursive self-improvement while also presenting that trajectory as dangerous. The central milestone is OpenAI’s claim that it has achieved what it previously called an “automated research intern.” The Decoder reports that OpenAI defines that as a system able to complete clearly scoped research tasks under human guidance, including some tasks that would take an experienced researcher several days. The evidence remains self-reported: The Decoder notes that OpenAI did not provide a detailed independent validation, saying only that the milestone was reached “according to our measurements.” OpenAI’s next stated target is more ambitious. By March 2028, according to The Decoder, the company wants to build a full automated AI researcher. OpenAI also says humans still set research priorities, evaluate results, and make decisions about scaling, pauses, and deployment. The reported usage figures show how far agents have moved from occasional assistant to daily research infrastructure. The Decoder says OpenAI’s median researcher now consumes more than $600 per day in inference at API prices, while the 90th percentile exceeds $7,000 per day. Median researcher token output has risen 124-fold since December 2025, and agent runtime has exceeded human working hours since June. As of mid-August, OpenAI’s research organization was running 3.1 agent workdays for every human workday. OpenAI is not presenting those metrics as a direct measure of research progress. The Decoder reports that the company calls them relatively easy to collect but difficult to interpret, because their relationship to actual research progress is uncertain. Experiments per researcher reached a record in August since tracking began in early 2025, but that increase coincided with a large increase in compute capacity. OpenAI also says overall progress may rise more slowly than individual usage metrics because the least automatable tasks become bottlenecks. The kind of work being delegated is also uneven. Using a taxonomy from Epoch AI, The Decoder reports, OpenAI found growth across categories of research work, with the largest gains in writing research and infrastructure code, providing technical assistance, and monitoring training runs. Higher-level planning decisions remain a small share of agent output. OpenAI also used an agentic classifier to measure whether agents completed assigned tasks, according to The Decoder. The analysis was limited to tasks with clearly measurable outcomes and grouped by estimated human time. From January to July, success rates improved across several difficulty levels; tasks estimated at under 15 minutes succeeded 86% of the time without intervention. The report also described continuing limits on autonomy. The counterweight is Pachocki’s warning. The Decoder reports that OpenAI’s chief scientist says no lab has a strong enough grasp of alignment, monitoring, and control to safely manage where these systems are heading. That makes the release unusual: OpenAI is documenting rapid internal adoption of AI agents in frontier-model research while simultaneously saying the control problem remains open. Who benefits: OpenAI researchers appear to benefit most directly from agents handling scoped coding, infrastructure, technical-help, and monitoring tasks. Teams with enough compute and clear task definitions may see similar leverage, though the provided sources do not quantify results outside OpenAI. Who's exposed: OpenAI is exposed to the control and monitoring risks highlighted by Pachocki, especially as it pushes toward a full automated AI researcher by March 2028. According to The Decoder, Pachocki warns that no lab has solved control, alignment, and monitoring well enough.