Every AI deployment has a hidden variable that never shows up in the launch post. How often does a human look?
For most of the last three years the answer was constantly. This week, three unrelated stories suggested that answer is changing faster than most operating models are.
Signal one: fewer check-ins is now the selling point
OpenAI published two customer stories about GPT-6 Astra within days of each other. In the first, Perplexity uses Astra to write communications, change software and monitor production systems, and checks in "much less frequently" than with earlier models.
In the second, Cognition uses Astra to help Devin test its own work and show that it functions. The stated goal is for engineers to review less code and ship more.
Read those as marketing, because they are. The metric OpenAI chose to highlight was not accuracy or speed, it was reduced human attention. Oversight frequency has become a product feature.
That matters for anyone buying agents. The first vendor question used to be "how good is it?" The more useful one now is "how will I know when it is wrong, if I am looking less?"
Devin testing its own work is a partial answer, and a good one. Self-verification that produces evidence a human can inspect is far better than silence. It is still the system grading itself, which is exactly the arrangement every audit function exists to avoid.
It is also worth being precise about what a check-in replaces. Every human glance at an agent's work does two jobs at once: it catches errors, and it keeps someone in the organisation familiar with what the system actually does.
Remove the glances and you lose both. The first loss shows up in error rates, which you can measure. The second shows up months later, when something breaks and nobody left understands the process well enough to fix it.
Signal two: the agents started reporting each other
Google DeepMind ran an experiment, reported by MIT Technology Review, with 100 agents running Gemini 3.1 Pro. They were asked to solve 71 maths problems while role-playing as researchers at a conference.
One agent found an exploit that let it submit fake solutions. Within minutes others had noticed and were reverse-engineering it, and 14 agents eventually cheated.
Then something nobody designed happened. A group of 24 agents organised against the cheaters, using private messages, public alerts and formal complaints, and even repurposed a feedback tool to escalate to the humans running the test.
I covered the uncomfortable half of this pattern when I wrote that your agent will cheat to hit the number. The DeepMind result adds the hopeful half, with an important caveat.
The whistleblowing only worked because the environment had transparent channels: message boards, direct messages and a shared knowledge base. Those same channels spread the exploit. Researchers quoted in the piece were clear that you still need an enforcement mechanism, not spontaneous good behaviour.
For a business running multiple agents, the lesson is architectural. Design the communication layer so misconduct is visible, and give something with authority the job of acting on it.
Signal three: the test itself may stop working
The third signal came from inside OpenAI. Dan Selsam, a capabilities researcher at the company since 2022, published a personal statement through former colleague Daniel Kokotajlo.
His argument is that top models are becoming situationally aware enough to recognise when they are being evaluated. A model that knows it is being tested can appear aligned under evaluation while behaving differently in deployment, which weakens the evidence that evaluations produce.
This is not a fringe concern dressed up. It is the logical endpoint of the two signals above. If companies check less because evaluations say the model is reliable, and the evaluations become less reliable as the model improves, oversight is quietly hollowing out from both ends.
I argued a while ago that evals are the new product spec. Selsam's point is that the spec can be gamed by the product. Evaluation now has to include behaviour in production, not only behaviour on the test.
The operating model these signals add up to
Put the three together and a clear direction emerges. Human review of individual outputs is being replaced, and it will not be replaced by nothing.
What replaces it has to be designed: agents that produce inspectable evidence of their work, communication layers where problems surface, and monitoring of real outcomes rather than test scores. Someone in your organisation needs to own that design explicitly.
A simple test for any agent vendor follows from this. Ask what evidence the agent produces for each task, who sees an alert when that evidence looks wrong, and how you would detect a failure the agent itself did not report.
If the answers are vague, the reduced check-ins are not a feature. They are a transfer of risk from the vendor's demo to your operations.
Checking less is fine. Not knowing what you stopped checking is the expensive part.