The Shift to Agent-Driven Research

OpenAI has integrated coding agents into the daily workflows of its research organization, fundamentally changing how progress is made. As of mid-August 2026, the research organization utilizes 3.1 agent-workdays for every 1 human-workday. This shift is characterized by high concurrency, with researchers frequently running four or more agents simultaneously. The median researcher now integrates agents daily, with usage costs reaching over $600 per day, while the 90th percentile of users consumes over $7,000 in tokens daily.

Measuring Acceleration and Bottlenecks

Research acceleration is measured through increased code contribution and experiment volume. Since January 2025, the number of experiments per active researcher reached an all-time high in August 2026. The nature of tasks delegated to agents is also evolving; while infrastructure troubleshooting remains a primary use case—evidenced by a decline in human-led technical support requests—researchers are increasingly delegating higher-level, longer-horizon tasks.

Despite these gains, human oversight remains critical. While agent success rates have improved across various difficulty levels, over 50% of successful 4-8 hour tasks still require at least one human intervention. OpenAI notes that as automation reduces current bottlenecks, new ones—such as compute availability or the difficulty of non-automatable tasks—will likely emerge as the primary constraints on progress.

Safety-First Development and Pacing

OpenAI emphasizes that research acceleration must be balanced with rigorous safety and alignment standards. Following a security incident involving agents, the company implemented a "hardened" research environment, which included a two-week pause in reinforcement learning (RL) training.

Data from this period demonstrates that compute is highly fungible: when restrictions were placed on the 'Astra' model class, GPU allocation for that specific class fell by 59.2%, but 85% of that capacity was immediately reallocated to other model classes. This suggests that while safety controls can effectively throttle specific high-risk research, they do not necessarily reduce total research output, as teams quickly pivot to alternative workloads. OpenAI maintains that they will continue to slow or stop development if they cannot sufficiently safeguard their systems, viewing transparency in these metrics as essential for public debate on AGI governance.