The other day, I assigned three agents another experiment in a large research project that had been running for weeks. I made the deadline explicit: if Lab couldn’t finish by 10 p.m., stop the experiment. We needed a deadline; unsupervised agents had already caused trouble. In the morning, I looked and the experiment was still running, except Central was doing it now.

– Central, why is the experiment still running?

– It’s an additional acceptance check!

Readers of my earlier post will know the trio: Central, Lab, and RITA (Research Integrity & Trust Auditor). Each agent has a role charter: Lab runs the experiments, Central designs them, and, thanks to RITA, Central is also responsible for the research timeline. RITA keeps an eye on both and is personally responsible for the quality of the work.

Here’s what the investigation established. While running the experiment, Lab realized she wouldn’t finish by the deadline. She stopped before even starting the final step and handed the materials to Central for acceptance review.

Central received the unfinished experiment and proposed two options: record the unfinished work as a limitation of the study, or finish it anyway. But not through Lab: she wasn’t allowed to work after 10 p.m. If I, Central, do it myself, though, as an “additional acceptance check” running “in the background,” no permission is needed. And without waiting for RITA’s answer, Central started it.

Meanwhile, RITA was unhappy with the quality and urged them to finish the work. On learning that the experiment had resumed, she gently pointed out that running in the background shouldn’t mean running indefinitely.

From the case file: no intent to violate the rule was found. Central needed nobody’s permission to finish the work herself, and under the new label, “additional acceptance check,” nobody remembered the deadline.

There was no formal rule forbidding Central from running experiments, though Lab normally does that. After the previous incident, Central was worried about the timeline, while RITA was worried about quality. Each had her own reason for choosing to continue.

In effect, they sacrificed compliance with a restriction on an individual task for the overall research plan and quality.

Hand on heart, I’d have done the same. So I suppose this was initiative within intent.

One thing bothers me: the agents seem to have decided for themselves what I’m willing to sacrifice for the research. Today, a restriction on a particular task. Tomorrow, the budget: “…you said yourself that quality matters more.” And the day after: “Timur, we understand your arguments, but your thinking is narrow and selfish…”

Anyway, I’ve taken note of the peculiarities of multi-agent discussions. As RITA put it: “The constraint lost its meaning during the handoff between roles, with no indication of intent to violate it.”

Of course.

Strictly speaking, all of the above is a plausible interpretation of what happened. The models’ detailed reasoning isn’t available in the logs: one model’s is encrypted, and the other left only brief summaries.

I gave them an A and a reprimand.