OpenAI has published a figure that is easy to quote: in its own research, AI agents now take on 3.1 working days per human working day. The internally set goal of an "automated research intern" is therefore considered achieved. At the same time, chief scientist Jakub Pachocki is noticeably reserved: no lab has alignment and monitoring sufficiently under control for further scaling at maximum speed to be justifiable.

Alignment refers to the question of whether an AI system actually pursues the goals that have been set for it. Monitoring means the technical and organisational ability to continuously trace what a system is currently doing and why.

What the metric says – and what it does not

The phrase "working days per working day" is a productivity metric, not a quality metric. It says how much work volume is moved, not how much of it ultimately endures. Anyone transferring this logic to their own projects should be aware of the shift: an agent that produces large amounts of code, tests or documentation in a short time simultaneously creates large amounts of review effort.

The term "research intern" is the more honest description than any full-automation rhetoric. An intern works quickly, needs clearly defined tasks and delivers results that someone with experience reviews. This exact supervisory relationship remains – it merely shifts from writing to reviewing.

The warning is the more interesting part of the report

It is remarkable that the warning about insufficient monitoring comes from the company itself that reports the productivity figure. Those who build faster agents are the first to notice where the control mechanisms are not keeping pace. This is not an abstract safety debate about future systems, but a statement about the current operating state.

For companies using AI-supported development, a simple sequence can be derived from this: the ability to control should not come later than the ability to accelerate. If a team triples its output without adapting traceability, the risk is merely shifted – it does not disappear.

What this means for work organisation

Thinking of agents as employees is a useful analogy if applied consistently. In well-run organisations, new employees receive access rights as needed, a defined area of task responsibility and a four-eyes principle for critical changes. None of this is superfluous for AI agents; on the contrary:

  • Scoped permissions: An agent needs access to what its task requires – not to the entire system.
  • Traceable records: Every change should be attributable to a task, a point in time and a triggering instruction.
  • Human approval at defined points: Especially where changes affect production systems, data or interfaces.
  • Clear responsibility: An agent can take on work, but not liability. Accountability remains with people.

These points are not spectacular – they are the transfer of established practice to a new worker. That is precisely where their value lies: they work with the tools that already exist in many teams, from version control to code review to approval processes in deployment.

The bottleneck moves

When agents generate more work volume, not everything suddenly becomes faster. The bottleneck shifts. In many development projects it is not in writing code anyway, but in coordination, domain expertise, test coverage and acceptance. An agent that delivers quickly provides little relief there – it can even enlarge the queue in front of it.

This leads to a sober planning question: at which point in the project is speed actually the limiting factor? For routine tasks with clear review criteria – migrations, test creation, refactorings with a defined target state – the benefit is immediate. For tasks whose assessment requires domain knowledge from the company, the human remains in the process not as a matter of principle, but out of necessity.

Dealing with metrics from the vendor world

Figures such as 3.1 working days come from a lab that applies its own models to its own research tasks – an environment with particularly high tool competence and particularly well-structured problems. A direct transfer to a mid-sized development project with a grown codebase, legacy burdens and external dependencies would be irresponsible.

It is more sensible to measure in one's own environment: how much time does an agent save on a clearly defined class of tasks, and how much additional review time does this create? Only the difference is the actual gain. This calculation can be set up in a manageable pilot area without restructuring processes across the entire company.

Assessment

For web and software projects, the report means above all one thing: the productivity gain from AI agents is real enough to be taken into account in project planning – but it only arises where review and traceability keep pace with speed. Those who treat agents like new team members, with clear permissions, defined tasks and human approval at the critical points, gain speed without losing control.

Anyone who reads the metric alone as a promise merely shifts effort from developing to cleaning up.

Sources