OpenAI's coding agents now autonomously complete nearly half of certain research tasks, with success rates reaching 46% for two-to-four-hour tasks and 35% for four-to-eight-hour tasks as of October 2026. This marks a substantial increase from March 2026, when agents completed only 23% and 13% of these tasks, respectively, indicating a rapid acceleration in AI-assisted research capabilities.
This progress, detailed in OpenAI's November 18 report, "Research acceleration: The jagged frontier," highlights a growing reliance on AI tools within the company's research workflows. The report combines internal activity measurements with staff surveys, painting a comprehensive picture of how AI agents are transforming research inputs, processes, and outputs.
While agents are proving highly effective for specific tasks, human oversight remains crucial. The findings suggest a future where AI and human researchers collaborate more intimately, redefining traditional R&D paradigms.
How Are AI Agents Reshaping Research Workflows?
AI agents are fundamentally changing how research is conducted at OpenAI by handling an increasing number of coding and experimental tasks, allowing human researchers to focus on higher-level conceptual work. This shift leads to more experiments and lines of code produced, but requires humans to maintain quality control and strategic direction. The median researcher's daily spend on agentic AI tools reached $2,400 in October 2026, a 62% increase from September.The report indicates a surge in agent usage, with researcher demand so high that the median respondent would trade 30% of team compute or 40% of team headcount to retain access to advanced coding agents. This underscores the perceived value of these tools in accelerating research timelines.
Despite increased automation, humans remain central to the research loop. Agents perform well in tasks like monitoring training runs (79% success rate) and answering technical questions (72%), but struggle with tasks requiring nuanced judgment, like writing complex code for inference stacks (25%) or building evaluation environments (29%). This illustrates the "jagged" nature of AI uplift, as explained in the report.
| Task Category | Agent Success Rate (Oct 2026) |
|---|---|
| Monitor Training Runs | 79% |
| Communicate | 69% |
| Analyze | 58% |
| Launch Training Runs | 35% |
| Build | 43% |
| Write Code for Inference/Serving Stacks | 25% |
Researchers interviewed noted that agents often exhibit poor "taste" in research ideas or produce "AI slop" in code, necessitating close human involvement. This highlights the need for a collaborative approach where humans guide and refine the agents' outputs to avoid wasted cycles and maintain quality.
OpenAI's Focus on Safety and Alignment
OpenAI has significantly increased its allocation of human effort towards Safety and Alignment research, with documented workweeks in this area increasing by 77% from January to September 2026. This heightened focus, partly spurred by events like the Hugging Face incident in July 2026, aims to address the critical technical and governance issues associated with increasingly powerful AI systems. This move aligns with broader industry efforts to responsibly scale AI, as explored in articles like OpenAI Disputes Claim It Violated California's New AI Safety Law.The company also reports progress on its internal capabilities index, OpenAI ECI, which has increased by 16.4 points per year since August 2025. This metric helps track advancements in model capabilities directly. OpenAI advocates for other AI labs to publicly report indicators of recursive self-improvement (RSI), the process where AI accelerates its own development, to foster transparency and inform public discussion about AI's capabilities, risks, and necessary safeguards. However, OpenAI clarifies that its reporting does not imply an intention to pursue full RSI.




