Oscar Crowe Artificial Intelligence OpenAI’s ‘Jagged Frontier’: AI Agents Do Half the Work, Humans Still Decide

OpenAI’s ‘Jagged Frontier’: AI Agents Do Half the Work, Humans Still Decide

OpenAI published a follow-up report on November 18, 2026 that quantifies how much of its own research its coding agents now complete without human intervention, and the answer is both impressive and uneven: agents succeeded on 46 percent of tasks estimated to require two to four hours of human work, up from 23 percent in March, but managed only 25 percent of the job when the task was writing code for inference and serving stacks. The report, titled Research acceleration: The jagged frontier, matters because it is the most detailed public measurement yet of AI systems contributing to AI research, and because it argues that labs should be required to report these indicators.

The Headline Numbers

The headline is the duration buckets. For tasks estimated at two to four hours of human effort, estimated success without intervention rose from 23 percent to 46 percent between March and October. For tasks estimated at four to eight hours, it rose from 13 percent to 35 percent. OpenAI is careful to note that researchers may have changed which tasks they delegate as the models improved, so the numbers mix capability gains with shifting choices.

Where Agents Win and Where They Stall

The task-level breakdown explains the jaggedness. From August 1 through October 18, 2026, the estimated success rate was 79 percent for monitoring training runs, 72 percent for answering technical questions, and 35 percent for launching training runs. Writing code for inference and serving stacks came in at 25 percent, and building evaluation, reinforcement-learning, and agent environments at 29 percent. Grouped by category, Communicate scored 69 percent, Analyze 58 percent, and Build 43 percent.

The pattern is consistent: agents handle watching, summarizing, and answering far better than they handle standing up new systems. That is why OpenAI calls the frontier jagged rather than smooth, and why the company expects lower-performing categories to become the bottleneck on overall uplift.

OpenAI sorted these tasks with the O*NET AI R&D taxonomy developed by Epoch AI, which groups research work into categories such as Decide, Design, Build, Run, Analyze, and Communicate. The company publishes partial identification bounds and confidence intervals alongside the estimates, and it warns that the updated methodology is not directly comparable to the figures it released in September.

The Money and the Workdays

Usage tells its own story. At API list prices, the median researcher's daily spend on agentic tools reached about 2,400 dollars in October, up 62 percent from 1,500 dollars in September. The 90th percentile reached about 23,000 dollars, up 105 percent from 11,000 dollars. OpenAI warns that the trend bundles more activity, shifts to pricier models, and changes in agent harnesses, so it is not a clean price index.

Measured in time rather than money, agent workdays per researcher rose from 0.5 to 5.7 per day, an elevenfold increase. On pretraining and posttraining teams, agent runtime averaged roughly eight workdays per researcher per calendar day. In scaling and infrastructure, the average was roughly four. By mid-October, more than two-thirds of tracked researchers averaged a daily peak of at least four agents running concurrently, and more than 10 percent averaged at least sixteen.

OpenAI flags a measurement wrinkle in the spending figures. An internal harness efficiency change deployed on August 21 depressed recorded token utilization, and the company reverted it starting September 10, after which token spend climbed rapidly again. The trend is real, but the exact level in any single month depends on how the harnesses were configured.

Safety Got the Biggest Shift

The report also tracks where researcher attention goes. Safety and Alignment saw the largest absolute increase in documented workweeks of any research area, up 77 percent from January through September, driven by new hires and existing researchers shifting focus. OpenAI measures an inflection point that coincides with the aftermath of the Hugging Face incident in July 2026 and the company's subsequent work on safety cases for frontier training.

That finding cuts against the simple story that AI research automation only accelerates capabilities. The same tooling that speeds up pretraining also pushes the people using it toward the safety work those capabilities require.

Why Humans Still Decide

The survey data shows how dependent researchers have become. In a survey of 101 research staff conducted from late August to early September, the median respondent said they would trade 40 percent of team headcount or 30 percent of team compute to keep frontier coding agents for three months. Asked the same question for January 2027, the median answer rose to 50 percent of team compute. Respondents also said their team would need a median of three months to produce work of equal value to one month of work if limited to GitHub Copilot and ChatGPT.

The origin of ideas tells a more grounded story. Posttraining and safety monitoring researchers said people came up with all the ideas they named, with models helping implement them. Pretraining researchers attributed the ideas behind GPT-6 Astra entirely to people. For GPT-6.2 Astra, pretraining researchers described developing ideas in conversation with Codex, with one crediting Codex for identifying key insights.

Interviews with ten researchers across pretraining, posttraining, and safety reinforced the same picture. They described agents that write slop-like pull requests, produce unusable data artifacts, and rabbit-hole on unnecessary steps. One posttraining researcher said the reliable pattern is to have an agent produce one example, review and edit it, and only then have it repeat the task. Several researchers expected the current bottlenecks to improve quickly, and one admitted he had no confidence his own taste in ideas would stay ahead of the models for long.

What the Outputs Show

OpenAI tracks three output measures. Lines of code changed per active contributor rose sharply through 2026. Logged experiments per active experimenter averaged 1.68 times the 2025 baseline over the four weeks ending October 18. And the company's internal version of the Epoch Capabilities Index, which mixes public and internal benchmarks, shows the OpenAI frontier rising 16.4 points per year since August 2025, compared with the 14 points per year Epoch observed across the field since late 2024.

Those are measures of activity and capability, not proof that AI is improving itself in a self-sustaining loop. OpenAI says as much. It argues that labs should track and publicly report indicators of recursive self-improvement so the public can debate the safeguards these trends require. The data in this report is the argument.

If you want more clear-eyed coverage of how AI models and AI businesses actually evolve, visit Crowe Endeavors.

Sources

"OpenAI maps the jagged frontier of AI agents speeding up research" (Trend Maxing (YouTube), November 19, 2026) — https://www.youtube.com/shorts/7uLu-exvFyQ

"OpenAI researchers use $2,400 of AI a day. Where it still fails" (Emergent Mind (YouTube), November 18, 2026) — https://www.youtube.com/watch?v=KbWQf9FFhro

"Research acceleration: The jagged frontier" (OpenAI, November 18, 2026) — https://openai.com/index/research-acceleration-jagged-frontier/

Leave a Reply

Your email address will not be published. Required fields are marked *

Related Post