On November 18, 2026, OpenAI published its second report on how coding agents are changing its own research organization: "Research acceleration: The jagged frontier." It follows up on the September report that claimed OpenAI had reached its "automated research intern" milestone. The new numbers are more interesting than the milestone — not because they are uniformly impressive, but because they are uneven in specific, measurable ways.

The headline: agents now complete nearly half of two-to-four-hour research tasks with zero human intervention. But the report's own framing word is "jagged," and the data backs that up.

The headline numbers

OpenAI's core metric is zero-intervention success — whether an agent finished a delegated task without a person stepping in. From March to October 2026:

  • Tasks estimated to take a human two to four hours: success rose from 23% to 46%
  • Tasks estimated to take four to eight hours: success rose from 13% to 35%

That is real progress, but OpenAI attaches an important caveat: as agents got more capable, researchers changed which tasks they delegated within each duration bucket. The improvement is a mix of better models and smarter delegation choices, and the report does not try to separate the two.

Where agents win — and where they stall

The task-level breakdown, covering August 1 to October 18, is where the "jagged frontier" becomes concrete. Estimated zero-intervention success rates:

  • Monitoring training runs for issues: 79%
  • Answering technical questions: 72%
  • Communicating, as a category: 69%
  • Analyzing, as a category: 58%
  • Building, as a category: 43%
  • Launching training runs: 35%
  • Building evaluation, reinforcement-learning, and agent environments: 29%
  • Writing code for inference and serving stacks: 25%

The pattern is clear. Agents are strong at watching, explaining, and analyzing — work where the cost of a mistake is low and the feedback loop is fast. They are much weaker at the load-bearing engineering that actually ships models: launching runs and writing the serving stack. OpenAI's own warning is blunt: as long as performance on AI R&D tasks stays heterogeneous, the weakest categories could become bottlenecks to overall uplift. A research org that can monitor 79% of runs autonomously but can only write serving code autonomously 25% of the time is not a self-driving lab; it is a lab with a very capable intern and a very human-shaped hole in the middle.

Usage is surging — and concentrating

The spending numbers tell a second story. At API list prices:

  • The median researcher now spends about $2,400 per day on agentic AI tools, up 62% from $1,500 in September
  • The 90th-percentile researcher spends about $23,000 per day, up 105% from $11,000 in September

Agent workdays per researcher — eight hours of agent runtime counted as one workday — rose from 0.5 to 5.7, an elevenfold increase. By mid-October, more than two-thirds of researchers averaged a daily peak of at least four agents working concurrently, and more than 10% averaged at least sixteen.

The gap between the median and the 90th percentile is widening fast. The heaviest users are not just using agents more; they are pulling further ahead of everyone else. For context, the API-equivalent value of agentic AI use in the research organization was about twice its non-equity compensation as of September 2026 — yet internal deployment still accounts for only 2% of total research compute. The agents are expensive in tokens and cheap in compute, relatively speaking.

Researchers would trade headcount to keep the agents

OpenAI surveyed 101 research staff with a deliberately stark hypothetical: suppose your team's access to all coding agents better than GPT-5.2 was revoked for three months — what would you trade to get it back? The median answers:

  • 30% of team compute, or
  • 40% of team headcount

In a separate late-October survey, 27 of 28 respondents said they expected to be more productive with OpenAI's latest internal model than with help from a mid-career human research engineer plus a 2024 model like o1-preview. These are self-reports from a small, ad-hoc sample, and OpenAI says so itself — but the direction is unmistakable.

The safety shift nobody is talking about

One of the report's quieter findings may matter most. Documented workweeks in Safety & Alignment grew 77% from January to September — the largest absolute increase of any research area at OpenAI — driven by new hires and existing researchers shifting focus. Growth picked up significantly after OpenAI's response to the Hugging Face incident in July 2026.

As agents take over more of the coding, humans are spending more of their time on the question of whether the whole thing stays under control. That is not a footnote; it is the shape of the job changing.

Why OpenAI is publishing this at all

The report repeats a policy ask from the September post: AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI) — the process by which AI accelerates the development of future AI systems. OpenAI stresses that publishing these indicators should not be read as an intention to pursue full RSI.

That framing matters. If agents can do nearly half of mid-length research tasks on their own, and the trend is still climbing, the question of who measures the acceleration — and who gets to see the numbers — stops being academic.

What to take from this

Three things stand out:

  1. The frontier is jagged, not smooth. 79% on monitoring, 25% on serving code. Any claim about "agents replacing researchers" that averages those numbers is hiding the interesting part.

  2. Delegation skill is now a research skill. The gains partly reflect researchers learning what to hand off. The bottleneck is shifting from model capability to human judgment about task selection.

  3. The measurement is the message. OpenAI is publishing internal metrics that no lab is required to publish, while arguing that all labs should be required to. Whether or not you trust the numbers — they are self-measured, with no independent audit — the fact that the field's leading lab is arguing for mandatory RSI reporting is itself the news.

The agents are not taking over AI research. They are taking over specific slices of it, quickly, while humans move to the slices that are harder to automate — and to the safety work that the automation itself makes more urgent.

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