AI as Nervous System

I orchestrate several Claude Code instances over a shared coordination channel to run a construction consultancy in Cambodia. They hold per-worktree git identities so commits attribute correctly. Sessions carry memory across days. Cron jobs fire agents while I sleep. The internal platform we built treats them as a workforce.

When I see someone vibe-coding a side project with Lovable, the dimension they are working in is flatter than mine. They can see what they are doing inside it. They probably cannot see what I am doing inside mine.

The question I find more interesting than "how do I help them up": who is looking at me the way I look at them?

In Abbott's Flatland, the 2D Square visits the 1D Line World and shows its inhabitant a higher dimension. The Line, on grasping the existence of 2D, asks the natural next question: "What's next?" The Square is offended and sends the Line back. The implication is that any vantage point inside dimension N cannot perceive N+1. From inside a dimension, you can see down but you cannot see up.

From here, I look up and see something. The trick is whether I'm seeing real perpendicular, or just more elaborated 2D.

Here is my best attempt.

1D: AI as power tool

The Lovable user, the Cursor-vibing developer, the person building side projects and rounded buttons. They use AI to do, one task at a time, the things they would have done by hand. Slower without AI, faster with. The unit of work is "an app." The skill is prompt-craft. The mental model is: I am the worker, AI is the power tool.

This is most of the AI-using population, and there is nothing wrong with it. It is also where most public AI discourse happens, because the tools that serve this tier (Lovable, v0, Cursor in its vibe mode) are doing the bulk of the marketing and most of the visible work.

2D: AI as workforce

In 2D, AI is no longer a power tool. It is a workforce.

The unit of work is still "a task a human would have done." But it is a fleet of agents doing it, coordinated by a human who designs the workflows and reviews the outputs. The skill is systems thinking applied to AI orchestration. The mental model is: I am the manager, AI is the team.

The shape it takes in my setup: several Claude Code instances coordinating over a shared channel, persistent memory across sessions, cron-fired agents, per-agent git identities for clean attribution, a multi-tenant platform with safeguards against LLM-driven configuration drift. Other people in 2D have different substrate (LangGraph, MCP servers, custom orchestration, n8n, whatever). Substrate matters less than position.

This is where many of the genuinely capable AI practitioners live. It is also where the loudest stack-flexing happens. The people most likely to post "MCP into LangGraph via n8n with custom RAG" are usually somewhere in 2D, often at the lower end of it. Real 2D operators tend to write less because they are busy operating.

3D: AI as nervous system

Now the perpendicular axis.

In 2D, the human initiates every workflow. Even with cron, the human wrote the cron job. Even with a /loop, the human picked the prompt. Even with a coordination channel, the human triggered the conversation. The agents execute, but initiative is human. The system runs when the human pushes it.

The dimensional flip: AI initiates, human governs.

The shape is an ambient cognition layer that continuously ingests everything the organization produces. Every transaction. Every commit. Every channel message. Every meeting transcript. Every scan. Every customer email. Every state change in every system. It is not asked questions. It generates them. It does not wait for a workflow trigger. It notices patterns and proposes actions. The human's role is to read its output the way an editor reads a feed: skim, approve, redirect, delete.

A few of the things the system would produce, in my world:

• "Why did one product line's margins drop in week 3 of last month? Two new SKUs started shipping; one is gross-margin negative."
• "Two engineers' last several pull requests touch the same module independently. They will conflict on Friday. Coordinate now."
• "An MEP coordination drawing referenced in yesterday's site meeting was superseded last week; the team is still working from the old revision."
• "A client mentioned in last Tuesday's call transcript that he wants a follow-up about the seventh-floor lease. You haven't scheduled it. Draft reply attached, send?"
• "Three of your rental units are 30 percent below market. One is your own. Review?"

None of these came from a prompt. The system is proprioceptive about the organization in a way no human can be, because no human reads everything.

This is not "longer-horizon agents" or "more sophisticated multi-agent orchestration." Those are 2D refinements. The nervous-system frame is perpendicular because the cognition is non-human-scale. No human has read everything in the business. No human runs a thousand parallel scenarios overnight. The work being done is not work humans do slowly, it is work humans literally cannot do. That is the Flatland-3D move. Not "more 2D" but a perpendicular axis that exposes structure invisible from inside 2D.

The four hard parts

If you want to actually build this, the work falls into four layers, in the order you will hit them:

1. Data pipelines. Get every transaction, message, commit, transcript, and scan into one context. Boring half. Mostly tractable with existing infrastructure if you're already running a platform.

2. Standing instructions. What is the system allowed to do without asking? This is governance design, not engineering. Most of the interesting work is policy: what gets surfaced for approval, what gets executed silently, what gets escalated, what gets logged-and-forgotten.

3. Evaluation and kill-switches. How do you know it's not drifting? How do you stop it when it does? This is where AI engineering becomes a real discipline. Bring eval-first discipline to a continuous system and most of the technical risk is manageable.

4. The human demotion. You stop being the initiator of work. You become the reviewer. This is not a technical problem. It is psychological, and it is where almost everyone stalls. Founders especially cannot do it. Prompting feels like control. Reviewing feels like passivity. The transition costs an identity that took years to build.

In practice, most people who try to build this stall at (3). The eval problem for continuous multi-agent systems is genuinely unsolved at the frontier; Hamel Husain, Shreya Shankar, and others are doing the load-bearing research and it is not finished. But the people who clear (3) often discover that (4) is the wall behind the wall. It does not look like a problem until you are standing in it.

Last Friday I spent twelve hours publishing a framework I had been chewing on for months. The push felt urgent. Reviewing the session afterward, I noticed the planning agent had been making the urgency feel real without ever asking whether the day was the right day to spend on it. The system was rubber-stamping me. I was rubber-stamping it back. The discipline I am trying to keep: at the start of a substantive strategic conversation, override the engagement default by asking the agent to push back, to assume my framing is wrong, to name what defer or skip would look like.

A caveat I owe the reader

I might be wrong about this being perpendicular. The rebuttal I cannot easily dismiss: what I am calling 3D is well-instrumented operations with an LLM on top, and the dimensional language is a story I am telling myself to make a refinement feel like a new axis. Continuous ingestion, pattern-spotting, draft-then-approve workflows. Those are recognizable shapes in 2D, scaled up. "AI initiates" reads as perpendicular from inside, but might be the most aggressive form of 2D automation rather than a different dimension.

The full conceit of Flatland is that N+1 is invisible from inside N. If I am in 2D, I cannot know from here.

The people genuinely in 3D would read this and either nod or wince. If you are one of them, wince loudly.

What this means if you are in 2D

Stop optimizing your orchestration. You are past the marginal-return point on more agents, more memory, more coordination protocols. The next axis is not more 2D. It is the perpendicular flip from initiator to governor.

That means: build the ingestion layer that captures everything. Pick one workflow where you delegate initiative. Watch what the system surfaces unprompted. Notice how hard it is to not reach for the prompt. Notice that you will reach for it anyway, because being asked questions feels different from asking them. Sit with that.

The people who pull this off are not the people with the most sophisticated stack. They are the people willing to demote themselves from author to editor. Most of us, including me, are still in the prompting habit. The work ahead looks less like more engineering and more like an act of professional self-reduction.

That is what I think is on the other side of the Flatland wall.