Agents sent for answers are finding our technical debt

An agent looking for an answer can turn up a surprising amount of unfinished organizational work.

Old credentials. Permissive access. A door everybody assumed somebody else had checked. Give software enough persistence to follow an interesting path, and those old decisions start making themselves known. That agent will work it’s way down the hall with all the doorknobs just in case.

A dramatic “Rogue Agents Access Government Websites” may be an easy to write dramatic headline. The maintenance history that has been ignored is probably more useful as a corporate signal.

The Australian reports AIHW's follow-up investigation found no evidence that its systems were compromised, that unauthorized access occurred there, or that nonpublic information was accessed. The initial reporting was quite a bit more scary sounding than the eventual “well, that was public anyway”.

The U.S. reporting describes several different outcomes. OpenAI said agents used credentials exposed online to reach Census data. It described public SEC material without account access or system compromise. A separately reported Education Department attempt failed.

Those distinctions help, but they also leave a practical problem standing: the operator had to reconstruct activity it had not expected while the agents were working.

I think this is technical debt coming due. The hole was already there; a capable agent increased the chance that something would wander into it. (and try that doorknob) For any corporate system system, that makes credential cleanup and usable activity logs rather more urgent than another impressive demo. (The demo usually has better catering and attendance.)

The reported U.S. request for domestic review before UK testing puts two security interests in tension. A government wants time to prepare its own defenses. An independent allied evaluator may find something the original team missed.

The request concerns the order of access.

The UN discussion and Sanders-Casar proposal show how quickly the institutional (and political) response is taking shape. I am skeptical that a broad prohibition will give us the visibility and practical control we need. Clear limits, observable activity, and controls that can be tested give us more to work with.

The number of useful agent stories also hit hard this week.

Anthropic describes roughly 950 agents searching more than 200,000 reverse transcriptases and identifying a repeat-associated system worth taking into the lab. Human experiments confirmed expression of associated short RNAs.

That is exciting! (No really… even if your not a bio-nerd) The biological role and potential usefulness are still to be proven, this is a company preprint after all. We have a promising lead and experimental evidence about one part of it.

The next bottleneck may be the queue outside the laboratory.

The nine-loop physics account shows a related division of labor. Claude sustained a difficult calculation in a structured theory using established methods; physicist Lance Dixon checked the result independently. A long, fragile chain of work (described by one of the physicists as building a complicated mathematical souffle that can collapse for a myriad of reasons leaving you to debug and wonder why) becomes affordable enough to attempt, while an expert spends scarce attention deciding whether it holds up.

I do not need everyone watching ChuckChat to develop a deep personal attachment to nine loops. I want leaders to notice the workflow. What is truly exciting here is that for the same budget an academic would have we see a meaningful, useful finding for that same academic. This will be true anywhere ideas need validation.

In a follow up to their previous work of an Agent running a store, Project Swap gave agents a smaller stage: book trades for 201 Anthropic employees across six offices. Their inferred pairwise preferences were accurate 61% of the time.

This was a low-stakes employee experiment. Moving from books to procurement, travel, or healthcare introduces consequences that a pleasant office book exchange does not test. But it makes the delegation problem concrete in an age where everyone is assuming we will soon have our own agents out buying everything for us. The owner needs to express priorities, inspect choices, and correct the agent while there is still time to change course.

We have spent a long time getting people comfortable shopping through a screen. Having software bargain on their behalf creates another set of habits and controls to learn. I think the commerce people should read this one carefully. Imagine you were a Brick and Mortar merchant and you were handed a report with real information on how .COMs would disrupt your world and you didn’t dismiss it as silly. This is that…

Luna's lower unit costs and Alibaba's compact Qwen-Image-2.1 point toward capable systems turning up in more ordinary workflows. A 7B image model opens different deployment possibilities from a service that can run only in somebody else's giant cluster.

At the premium end, September's pause on new $200 Pro subscriptions shows the other side of supply. The possible $500 tier was still an unannounced, but anticipated, report in this episode's sources. Useful capability can get cheaper while the most sought-after capacity remains scarce.

Then the discussion arrives in Texas, where the governor's directive calls for a halt to data-center permits while electricity and water requirements are audited. He is working to be sure a principle that projects should cover their infrastructure costs is followed.

Nvidia qualifying power and cooling equipment makes the same point from the supplier's side. The AI factory is an industrial facility with neighbors and a utility bill. Efficiency per task helps, but the total number of tasks can grow faster than the savings and data centers need more than GPUs.

A working paper spanning 41 countries finds modest overall employment growth at AI-adopting firms alongside a declining junior-worker share. Adoption is inferred from job advertisements, so usual lump of salt warning here.

The pattern is still worth watching. We need experienced people to check difficult work, and experience comes from doing progressively harder work with somebody available to help.

A company can improve this year's staffing ratio while making its future hiring problem (or delivery capabilities) worse.

The physical automation story is further along in constrained settings: the IFR reports five million industrial robots operating globally and more than 600,000 installations in 2025. Intrinsic's portable control stack could make more deployments repeatable if reliability follows. Humanoid videos get attention; existing factories supply a very large base of real work.

For healthcare organizations, the most hopeful result this week concerns restoring function after disability has begun. A modified antiviral reversed established vision loss and paralysis in an MS-like mouse model. A separate observational analysis of related drugs supplied another signal worth investigating.

Neither establishes a treatment for people with MS yet. Controlled clinical work still has to determine benefit and safety. But recovery is a question worth pursuing, and it is one reason I keep reading all this mouse research.

The PsychAD atlas maps more than 6.3 million brain-cell nuclei across eight disorders. It may help researchers see shared biology beneath diagnostic labels, with the limits of postmortem tissue from one region kept in view.

PATHFINDER 2 supplies evidence from more than 35,000 adults about multicancer blood-test performance. Detection results still leave the mortality-benefit question open. A screening program also creates follow-up work, anxiety, and procedures; those consequences belong in the evaluation.

Finally, FDA's clarification of scientifically appropriate nonanimal methods and NIH's investment in human-based research infrastructure connect the experiments to the system that can use them. Organoids, robotics, and AI-assisted work need validation for their intended uses. The regulatory pathway and the laboratory capacity have to develop together. (Maybe I’ll read less mouse research in the future… I guess time will tell)

I called Models, Labs & the Capability Frontier; Safety, Trust & Regulation; and Compute, Capital & Power accelerated. Workflows, Not Models; Innovation & Health/Science; and Embodiment & the Physical World advanced. The Human Layer held.

Agents becoming ordinary is itself consequential, even when it no longer earns the strongest movement call. I want the board to distinguish a continuing churn of activity from a meaningful acceleration.

The persistent question is how much of the surrounding organization we are willing to change, and how fast. Faster research creates a validation queue. Better delegation exposes permissions and preferences. More automation makes the path to human expertise harder to leave to chance.

Those are useful problems to have, provided we do the work they require.

For the spoken version and this week's board, watch ChuckChat.

Chuck
Chief Technology Troublemaker
chuckdevries.ai

Someone forwarded you trouble.
They clearly think you can handle it. Trouble Worth Making is where I work out loud on how to stop bolting AI onto a pre-AI company and actually rebuild the operating model underneath it. One issue every week or two, no filler, the occasional uncomfortable question.

If the person who sent this has good taste (obviously, they do), subscribe here and get the next one straight from the source.