# I Automated So Much Work I Had to Invent Management Again

> I complained that I could not keep up with my agents. Then we looked at the private Git history. Apparently I had been documenting the birth of robot middle management for months.

The Git history tells the story: every time AI removed one bottleneck, the next layer of coordination became the problem.

**Published:** 2026-10-06  
**Tags:** AI agents, operations, Git, management, human agency, process  
**Canonical URL:** https://blog.nikdesign.ca/posts/i-automated-so-much-work-i-had-to-invent-management-again

---
This article exists because, while AI and I were writing another article about AI, I complained that I could no longer keep up with all the AI agents working for me.

That sentence is already ridiculous.

It gets worse.

I said there were repos to manage other repos. Agents to manage other agents. Systems to tell the agents which systems were authoritative. A mission-control repo whose job was increasingly to explain where everything else lived.

Then I said the obvious thing:

**Necessity is the fucking mother of all inventions.**

So we did something useful.

We checked Git.

And Git had receipts.

The repositories behind this story are private. They contain internal tooling, operating instructions, experiments, and company context. They do not need to be public for the story to be public. I used their creation dates and commit history as the primary-source record of what I built, when I built it, and what problem I was apparently trying to solve next.

## June: the first management layer

On June 3, I created a repository whose job was essentially to help maintain repositories.

Five days later, another operational layer appeared.

By June 14, even the Markdown and context accumulating around the work needed its own management surface.

I did not sit down with a whiteboard and announce that I was designing an organizational architecture for autonomous software.

I was annoyed by things.

So I built things.

The pattern was already there.

I just did not know it was a pattern yet.

## September: oh fuck

By late September, the problem had escaped the individual repository.

One internal audit covered **57 active, maintained, infrastructure, or incubating repositories**.

Another operation rolled repository-local tracking scaffolds across **27 active repositories**.

And then came one of my favourite moments in the fossil record: the tooling had become complicated enough that an agent had to investigate whether one repository-management tool duplicated another repository-management tool.

Read that sentence again.

**We needed tooling analysis to determine whether the tools for managing the tooling were overlapping.**

At some point the joke becomes architecture.

## October: the organization appears

October is where the commit messages stop looking like ordinary software development and start sounding like somebody is writing a constitution for a tiny machine civilization.

The records start saying things equivalent to:

**design the repository initializer**

**add a portable repository initializer**

**propose a shared repository manifest**

**make one system the single agent front door**

**define mission control as the capability and repository router**

That progression matters.

Agents could already do the work.

The emerging problem was that an agent entering the system needed to know:

Where am I?

What is this repository?

Which instructions are authoritative?

Which tools are already adopted?

Where does work get tracked?

What can I change?

What should I leave alone?

Where do I go next?

In a human company, we would call a lot of that onboarding, policy, management, governance, institutional knowledge, and organizational design.

Apparently machines need some of it too.

## Then I needed a manager

On October 4, a new repository appeared for standing agents.

Its early history is almost too perfect.

The commits define integration boundaries. Then they compose the canonical systems. Then they route a Shift Manager through those systems. Then they add an action console. Then they add a deterministic shift report.

And almost immediately the repository starts filling with reports and handoffs.

Why?

Because I had reached the stupidly obvious limit nobody mentions when they show you a demo of five agents running simultaneously:

**I am still one person.**

The agents do not get tired.

My attention does.

So execution stopped being the bottleneck and **management became the bottleneck**.

Which is how I managed to automate enough work that I had to invent management again.

## And then the agents needed sidecars

The next day, October 5, another tool appeared: a dedicated operating surface around a coding agent.

Within hours its history moves from an initial console prototype to documentation, standards alignment, session organization, an agent chat panel, and a desktop shell.

October 6: another agent gets its own sidecar.

Of course it fucking does.

None of this was the plan.

That is the important part.

## The architecture was caused by the work

There is a temptation in AI right now to draw magnificent agent architectures before the agents have actually done anything.

Orchestrators.

Supervisors.

Memory layers.

Agent registries.

Governance planes.

Swarm diagrams that look like somebody dropped a plate of spaghetti on a systems-design interview.

Some of those abstractions will be necessary.

But I have become suspicious of designing them before the operational pressure exists.

My own system evolved in the opposite direction:

**pain → tool → more capability → new pain → next abstraction**

The previous solution created enough leverage to expose the next constraint.

Repository maintenance became portfolio management.

Portfolio management became routing.

Routing became standardization.

Standardization enabled more agents.

More agents created an attention problem.

The attention problem created management.

That is not a roadmap.

It is evolution.

## The commits are the story

This is why looking at the Git history was so useful.

Memory tells stories backward. Once a system exists, it is easy to imagine that you always knew where it was going.

Git is less charitable.

Git says: no, motherfucker, on this date you were solving **this** problem.

Then three days later you discovered **that** problem.

Then you created another repository because apparently that was easier than accepting peace.

The commit history preserves the sequence before hindsight gets a chance to clean it up.

That makes it more than engineering provenance.

It is a record of pressure.

## Also: I did not type this

There is one more layer of absurdity worth disclosing.

I have a [Process page](/process) on this blog because I do not want the role of AI in the writing to be some coy little secret.

But even saying "AI helps me write" does not quite describe what happens.

**I do not type these blog posts.**

I talk.

Speech becomes text.

The conversation branches. I tell stories. I make arguments. The AI pushes back, researches things, checks repositories, finds connections, and helps shape the material. I decide what matters. We turn the conversation into an artifact. The artifact goes through Git. The site deploys.

For this post, the chain was basically:

**thought → voice → transcription → dialogue → Git archaeology → synthesis → artifact → Git again → deployment**

Yes.

We used Git history to write a story about Git history and then put the story back into Git.

I am aware of how stupidly recursive this has become.

## The interface is disappearing

But that recursion points at something bigger.

I am doing less and less work at the level of the traditional computer interface.

I am not manually opening every repository and inspecting every file.

Agents do that.

I am not manually maintaining every status surface.

Agents increasingly do that.

I am not manually typing the prose describing what happened.

I am talking.

The work is moving upward.

From keystrokes to instructions.

From files to intent.

From execution to judgment.

From doing every operation myself to deciding what deserves attention.

That does not remove the human.

It changes where the human sits.

And right now the scarce resource in my system is becoming painfully obvious.

It is not compute.

It is not typing speed.

It is not how quickly another agent can generate code.

**It is my attention.**

So the next invention will probably be whatever makes that problem hurt less.

I have no idea what it will be yet.

Give it a few fucking days.