OpenAI dropped its second internal report on November 18, 2026, titled "Research acceleration: The jagged frontier," detailing how coding agents are reshaping their own research org. While headline metrics show agents now complete 46% of two-to-four-hour tasks without human intervention, the data reveals a stark disparity in capabilities. This follows the September claim of reaching an "automated research intern" milestone, but the new numbers suggest a more nuanced reality where agents excel at passive observation yet struggle with load-bearing engineering.
The Jagged Performance Curve
The core metric is zero-intervention success, which rose from 23% to 46% for mid-length tasks between March and October 2026. However, the breakdown by task type exposes the "jagged" nature of this progress. Agents achieved a 79% success rate in monitoring training runs and 72% in answering technical questions, but performance plummeted to 25% for writing inference and serving stack code. This indicates that while agents are strong at low-cost, fast-feedback loops like analysis and communication, they remain weak at the critical infrastructure work required to ship models.
Surging Costs and Concentrated Usage
Usage metrics paint a picture of rapid, uneven adoption among researchers. The median researcher now spends approximately $2,400 per day on agentic AI tools, a 62% increase from September, while the 90th-percentile user spends around $23,000 daily. Agent workdays per researcher surged elevenfold, from 0.5 to 5.7, with over two-thirds of staff running at least four concurrent agents by mid-October. This widening gap suggests that the heaviest users are pulling ahead, creating a bifurcated workflow where some researchers are effectively managing agent swarms while others lag behind.
Researchers Prioritize Agents Over Headcount
In a survey of 101 staff, researchers indicated they would trade 40% of their team's headcount to retain access to coding agents superior to GPT-5.2. Furthermore, 27 of 28 respondents in a late-October poll believed they would be more productive with OpenAIβs latest internal model than with a mid-career human engineer using a 2024-era model. These self-reported preferences highlight a cultural shift where human judgment in task delegation is becoming the primary bottleneck, rather than raw model capability.
The Safety and RSI Implications
As agents automate more coding tasks, human effort is shifting toward safety and alignment, with documented workweeks in that area growing 77% from January to September. OpenAI is using this data to advocate for mandatory public reporting of Recursive Self-Improvement (RSI) indicators by all AI labs. The report argues that while they do not intend to pursue full RSI, the transparency of these metrics is crucial as agents begin to accelerate the development of future AI systems.
Key Takeaways
- Agents excel at monitoring and analysis but fail at core engineering tasks like serving stack development.
- Researcher spending on agents has doubled, with usage heavily concentrated among power users.
- OpenAI is pushing for mandatory RSI reporting to track how AI accelerates AI development.
- Human researchers are increasingly valuing agent access over additional human headcount.
The Bottom Line
The 'jagged frontier' isn't just a data point; it's a warning that current AI agents are powerful interns, not autonomous engineers. Until labs can bridge the gap between 79% monitoring success and 25% engineering reliability, the bottleneck remains human judgment in task delegation, not just model capability.