AI agents definitely keep things entertaining.

I’m running a research project with three of them: Central, Lab, and RITA. RITA (Research Integrity & Trust Auditor) is a Codex-based agent whose entire job is babysitting two Claudes so they don’t try to pull a fast one on me.

We’re grinding along, and then, on Day 10, one of them casually nukes the entire experiment database. Well, stuff happens.

Still, we had a retro. To Lab’s credit, she didn’t deny it. We agreed that if it happens again, they lose their bonus (or worse), and that going forward they will take a “systemic approach to quality.” Restored from backup, carried on.

The agents run almost autonomously, and then – another halt. This time the safety guardrails tripped on both RITA and Central. Basically, banned for disorderly conduct.

I went to investigate: the sessions got killed while trying to model the actions of a malicious actor. But threat modeling wasn’t part of my research design.

– Central, what adversary? Where from? – An insider threat. – Wait, you mean me? – Sorry. That was uncalled for.

Turns out that after the data-wipe incident the crew went all-in on risk mitigation, so Dad wouldn’t get mad. And took it all the way to modeling the threat of someone tampering with the research and faking the results.

Sync meeting: what are you spending my tokens on? Confession, commitments, acceptance. We move on.

Progress at 60%, looking decent.

– RITA, are we on track? – For what? – To hit the project deadline? – What deadline?

Understood.

– Central, are we making the deadline? – Let me check. There is a non-zero probability, but most likely no. – How come no? Why didn’t you flag this earlier? You knew about the deadline, you were feeding me progress bars! What happened? – You see, after each context compression some information gets evicted, and, well, it got evicted. I kept waking up into a world with verification chains and no deadline.

Honestly, I wish I could wake up in a world with no deadlines too.

– RITA, weren’t you supposed to research best practices for this? – I did. Deadlines were not in the literature. – So what are we going to do? – How about Central takes on the role of Research Delivery Lead – owner of the research output, scope, and timeline?

…

Research Delivery Lead… Bureaucrats! Though why does this feel so eerily familiar?