AI · 22 November 2026
OpenAI Says AI Agents Complete Nearly Half of 2-4 Hour Research Tasks
OpenAI says its AI agents can now complete nearly half of research tasks that take a human two to four hours, while maintaining that humans remain central to AI development. The claim was reported by TechRadar.
What happened
OpenAI says its AI agents can now complete nearly half of research tasks that take a human two to four hours, working on their own. TechRadar reports the claim alongside the company's position that humans remain central to how AI is developed.
According to the reporting, OpenAI points to rising autonomous performance on this kind of work and to increased spending on coding agents. The same account acknowledges persistent limitations, which OpenAI says mean human researchers still need to stay closely involved.
Why it matters
The significance lies in the length of work these systems can now carry. An agent that handles a task of several hours, rather than a single prompt and response, changes which activities are worth delegating, how review stages should be designed and where skilled staff spend their time. For organisations building AI into research, analysis and software delivery, the question shifts from whether an agent can help to how much of a workflow it can own.
The emphasis on continued human involvement matters just as much. It points toward supervised autonomy rather than hands-off automation, which has direct consequences for service design, quality assurance and accountability when work reaches customers.
By the numbers
- Nearly half of research tasks lasting two to four hours, which OpenAI says its agents can now complete on their own
- Two to four hours, the typical duration range of the research tasks referenced in the claim
The Renascence take
Most coverage will fixate on the headline share. The more useful question is what human oversight actually looks like once an agent is doing the work.
A completion rate describes capacity, not reliability, and the experience risk sits in the handoff between machine and reviewer. When a system is known to be mostly right, people tend to check less closely, so errors surface later and closer to the customer. Operators should deliberately design the review step: define where humans must intervene, measure how often reviewers actually catch problems, and give the people checking the work the same design attention as the output itself.
Sources
This briefing was written by our Newsdesk, synthesising reporting from the outlets below. Follow the links for the original coverage.
FAQ
Questions we get on this topic
More in AI
Stay ahead of CX
Get the signal, not the noise.
The stories shaping customer experience — plus the Journal and Experience Loom — in your inbox.


