The New Mouse
This post assumes that you already are familiar with the Country/City Mice dichotomy, which I had expanded into a third one called Cloud Mouse here:
And much before mine, there was a Metro/Cloud mouse dichotomy that Venkatesh Rao had written (though — you got to trust me when I say I came up with the Cloud Mouse terminology independently).
As a TLDR: Country Mice are people who are more “rural/unsophisticated” in behavior; City Mice are people who are more “urban/sophisticated” in behavior; Cloud Mice are people who bring that sophistication into virtual interactions — As in “Cloud” arena. (There are nuances, so please read the link above if your brain is raising exceptions with these one-liners).
However: We are fast moving into a new world, where the mouse needs to learn new levels of sophistication. Just sophistication with dressing/manners (like City Mouse) or the sophistication with virtual interactions (like Cloud Mouse) isn’t enough. The new-age mouse needs to be comfortable dealing with non-humans, starting with the virtual world, but after a while in physical spaces too. Your guess is right — I am talking about the interactions with non-human agents through LxMs. (LLMs, LAMs, LIMs, LWMs, and so on). Cloud sophistication optimizes interaction; GenAI demands orchestration.
Success is now dependent on how well the mouse can orchestrate the various agents to get the job done. Even if one doesn’t become a true Cloud mouse, they could still get by in the Cloud-driven world. Just that their work might be slightly slower, they may appear awkward in conversations etc.
But in the new world of Gen. AI & agents, it gets to the core of even how the job gets done. If the mouse doesn’t build expertise in dealing with agents, the output will not just be slow, but its quality will also be poorer than the work created by a mouse with the said expertise. This is a shift from working in systems (Cloud) to making them work (GenAI).
So, what shall we call this mouse? I did an elaborate brainstorming on this with… who else but ChatGPT. We discussed extending the “location” lens — from Country to City to Cloud to where next. But then realized that this one is not about “location” but transcends it. We have moved from where you live to how you behave to where you operate to finally what you control.
Get carnivas’s stories in your inbox
Join Medium for free to get updates from this writer.
There were several names that appeared nice (like “Vibe” mouse) but didn’t capture the essence of the skill well enough — it could end up a frivolous name that no one but Karpathy-fans understand after a few years. And then “mesh”, “synth”, “cyborg” and so on. Nothing seemed to do justice to it.
Finally, we chose to focus on the “skill” alone — of being able to work with AI systems and get things done through them. And to continue with it being a “C” word (like Country, City, Cloud…) we first came up with “Companion Mouse” (of an agent). Now, “companion” could mean a host of things and can also mean becoming a “pet” (thankfully not a paper-clip) of AI. So, to retain the human agency in it, the final name is…. drum roll….. “Conductor Mouse” (of many agents). While these are early days and things will evolve, as of now, I see a conductor doing some/all of these: conducting evals; juggling multiple agents; decomposing a task; correcting hallucinations; sequencing outputs; and so on.
It is a mistake to assume Conductor mouse as a higher-order Cloud mouse. It isn’t; instead, I would say it’s orthogonal. Some Cloud skills may transfer, and some may even be detrimental as is. In any case, some entirely new skills need to be developed.
Let’s now see an example of where Cloud mice will come up short without developing Conductor mouse skills. Consider the task of “writing a strategy document”.
A good Cloud Mouse is excellent at these: structuring arguments clearly; searching, synthesizing, and citing; writing polished documents; collaborating smoothly over docs, Slack, comments; and in general sounding coherent and “smart” in async conversations. Like I said earlier, even a City mouse would do all these, maybe a bit slow, that’s all. Once you give the Cloud mouse ChatGPT, they will use it to get a quick draft, tweak its tone, polish the language to suit their/their org’s style, and at most ask for a summary or counter-argument. The output is clean, fluent, and socially acceptable. This was very valuable in the Cloud era. All good, right? No. Cloud Mouse behavior becomes insufficient in these ways: the Cloud Mouse will likely ask one big prompt instead of decomposing the problem, treat ChatGPT like a better Google or junior writer, accept the first plausible-looking answer and edit around hallucinations instead of redesigning the process. The output looks good — but the thinking underneath is shallow, brittle, or subtly wrong. This is where Cloud skills plateau.
With Conductor mouse skills, this is what would happen: They break the task into sub-problems (facts, assumptions, options, risks), spin up multiple agents or prompts for different roles (analyst, skeptic, explainer, domain expert), explicitly ask one agent to attack the output of another iterate on the process (not just the prose) and discard entire drafts without emotional attachment. They are less impressed by fluency. In fact, they are suspicious of it. The final document may even look less elegant at first glance — but the reasoning is deeper, the tradeoffs clearer, and the blind spots surfaced.
This is why Cloud mouse behavior is party necessary, but definitely insufficient in the new age. And Conductor skills are learned over time. An overconfident Cloud mouse (think MBA from top colleges) will try to overfit some and get into the failure modes of Conductors. It could involve over-prompting, under-verification, mistaking fluency for truth, or just thinking orchestration means “more agents” rather than better structure or human thinking. The key is recognizing that being a Conductor isn’t about using more AI. It’s about using it more strategically. We could do an entire post on the failure modes, I think.
Edit (on 19-Jan-2026): I met the man himself yesterday (Venkatesh Rao — see top of this page) and asked him what could be some things a Conductor Mouse needs to be skilled at. Among other things he said, one thing that stood out to me was this — and I am paraphrasing:
In terms of the OODA loop, AI might be able to do the Observe, Decide, Act by itself. But it might not be able to Orient (or Reorient) as well as humans.
So that could be a thing for Conductor Mice to learn.
Now, without finding fault in all these arguments, go learn the skills to be a good “Conductor Mouse”. Don’t get stuck with being just a “Cloud Mouse”. That’s so 2022.
