Can a coding agent be given research work and expected to deliver a finished result? It depends on the specific task: in OpenAI’s data from August 1 to October 18, 2026, agents working without human intervention performed best at monitoring training runs, and noticeably worse at writing inference and model-serving code.
In its November 18 report, the company compared the tasks researchers assigned to agents and described how the use of these tools changed within OpenAI. The main practical takeaway is to treat “agent work” as a set of different tasks with different acceptance criteria. An average based on hours, or the impression that a model “writes code well,” is no substitute for checking a specific result.
What changed in the assessment
OpenAI’s previous report on research acceleration came out in September. The new version broadens the picture: the company added data on agent use, employee surveys, workflows, and success rates by task type. The results describe OpenAI’s internal research organization, however, and are not a universal assessment of all models or companies.
One of the clearest comparisons is the share of tasks agents completed without human intervention. For tasks estimated to require two to four hours of human work, the rate rose from 23% in March to 46% in October. For the four-to-eight-hour category, it went from 13% to 35%. These categories show agents’ ability to complete individual assignments without intervention; they do not indicate the share of all research work that can be handed over to a model.
There is also an important methodological detail: as agents’ capabilities changed, so did the choice of tasks assigned to them. OpenAI explicitly notes that rising success rates within time categories may reflect both model improvements and changes in the mix of delegated tasks. So the time-based comparison is useful as an internal trend, but by itself it cannot separate the model’s contribution from researchers’ decisions about which tasks to assign.
Why the task label matters more than its duration
In a snapshot of specific types of work, OpenAI reports success without human intervention of 79% for monitoring training runs, 35% for launching those runs, and 25% for writing inference and serving-stack code. Even two closely related operations around model training had sharply different results.


This explains why a broad category like “help with research” is a poor basis for planning. Monitoring a run to spot a problem is a task of observation and alerting. Launching a training run means correctly preparing the actions and execution conditions. Inference and serving code belongs to a different technical area. Combining these tasks into one metric would hide differences that matter in real workflows.
The company also reports higher rates for “communication” (69%) and “analysis” (58%) than for “development” (43%). But even these broad categories conceal differences between tasks. To choose a suitable assignment, it is more useful to specify a verifiable outcome—for example, detecting a predefined failure in a run log—and separately define what a human needs to confirm.
Heavy use does not mean equal returns
The report shows that agents have become a visible part of OpenAI’s internal workflows. Using API prices, the company estimates that the median researcher used agent tools worth about $2,400 per day in October; the figure for the 90th percentile was about $23,000. In both cases, these are equivalents based on published API rates, not figures for actual costs or evidence of a proven amount of budget saved.
Another measure is “agent-days”: every eight hours of agent use counts as one workday. On average, the figure rose from 0.5 to 5.7 agent-days per researcher per day. In pretraining and post-training teams, agents’ average combined work amounted to about eight such days per researcher per calendar day. Because several agents can work in parallel, the metric describes their total time, not the length of one employee’s workday; OpenAI also notes that it mainly captures interactive use and may undercount automation.
Increased use shows that employees are actively incorporating agents into their workflows. That alone is not enough to assess the benefits for another team: budgets, chosen models, agent-session lengths, and task types may differ. OpenAI’s data are especially useful as an example of which work metrics can be collected, not as a ready-made benchmark for every research group.
Where human work remains
In interviews with ten OpenAI researchers, the company described the practical side of delegation: agents helped write and debug code, analyze experiments, and review large datasets. At the same time, employees reported low-quality artifacts, incorrect assumptions in analyses, and inefficient attempts to launch experiments. The company also notes that employees attributed the ideas they named for GPT‑6 Astra releases to people, while models helped with implementation.
One approach to working with agents described by a researcher is particularly easy to adopt in a team: ask the agent to produce one example, check and correct it, then assign a similar series of tasks. This quickly exposes errors in the instructions or data format before they spread across the full set. For tasks that are expensive to verify, this is often more useful than scaling up the assignment immediately.
How to apply the findings to your team
Start with one repeatable task that has an observable result and a clear acceptance condition. Track how often intervention was needed, what errors occurred, and how long verification took. This information can help distinguish a task an agent genuinely takes off your hands from one where it merely speeds up draft work.
For high-consequence operations, set a separate boundary on authority. For example, an agent might detect an anomaly or prepare a launch command, while confirmation and execution remain with the responsible engineer. This is not a universal safety scheme, but a way to match the level of autonomy to the specific risk and available verification criteria.
OpenAI’s report offers a useful distinction: delegate specific assignments with verifiable outputs, not “research” or “development” in the abstract. In the company’s data, monitoring runs appeared better suited to autonomous completion than launching training or writing inference code. For a team applying these findings, the next step is to measure the same differences on its own tasks and expand delegation where results stand up to scrutiny.