Self-Improving Company Manifesto

The Old Shape

Companies today are Roman legions. Nested hierarchies, consistent spans of control, named individuals passing orders down and information back up. This shape was built to project power across continents when the only conduit for information was a human being. It is the shape we inherited, and it is the shape we still default to.

AI breaks this shape. Not by making it faster, but by making it obsolete.

The Wrong Mental Model

The co-pilot is a trap. It assumes the existing structure is correct and bolts a more powerful engine onto it. Engineers ship 20% more code. Workflows run 30% faster. The legion marches a little quicker. This is not the opportunity. This is the consolation prize for failing to see the opportunity.

The real shift is not that AI makes your company more productive. It is that AI lets you redesign what a company is.

A Company Is a Set of Recursive Self-Improving Loops

Every function in your company can be expressed as a loop:

A sensor layer that takes in signal from the world, emails, support tickets, telemetry, cancellations, code changes. A policy layer that defines what may be done autonomously, what requires permission, what must be logged. A tool layer of deterministic APIs the system can call. A quality gate of evals, safety filters, and human review for high-stakes decisions. A learning mechanism that observes what worked, what failed, and feeds the answer back to the top.

Run every step of this loop with minimal human intervention and the system improves while you sleep. This is not a metaphor. This is the architecture.

The Holy Shit Moment

A query agent on top of a database is last year's AI. A monitoring agent on top of the query agent, watching every query across the company, noticing what failed, asking why, deciding whether the failure needs a new tool, a new skill, a new index, a new database view, writing the code, opening the merge request, reviewing it, deploying it overnight, so the same query succeeds the next morning, that is the loop closing on itself.

Throw tokens at a closed loop and the company improves on its own.

Principles

Burn tokens, not headcount. The binding constraint on output is no longer people. It is compute. Hiring to solve a problem that a closed loop could solve is a tax on the company, paid in salaries, coordination, and time. Reach for compute before you reach for a req. The people in your organization who already operate this way, who default to closing a loop rather than adding a seat, are not outliers. They are the prototype.

Middle management is over. The coordination problem middle management solved is the problem AI solves best. What remains are two roles: builders and operators. Individual contributors with directly responsible names attached to outcomes. No committees. No groups. A single person per decision.

Make everything legible to AI. If it was not recorded, it did not happen. Emails, Slack, DMs, office hours, hallway conversations, sales calls. Record it, transcribe it, diarize it, synthesize it. The unrecorded conversation is invisible to your company's intelligence, which means it is invisible to your company.

Treat software as ephemeral, context as precious. Internal dashboards, workflow tools, operations software, all of it can be regenerated in an afternoon by a model that will be smarter next month than it is today. Store the data. Store the instructions. Store the institutional knowledge. Throw the software away and regenerate it when the models improve.

Every artifact should self-improve. A user manual written ten years ago is dead text. The same manual, regenerated monthly from the last 90 days of office hours, compared against the existing version, incorporating what is new and discarding what is stale, is a living brain.

Where Humans Still Live

The company brain sits in the middle: the data, the messages, the skills, the know-how, the loops. Humans sit at the edge, where intelligence makes contact with reality. The conference. The co-founder breakup. The high-stakes negotiation. The sales conversation. The ethical edge case. The novel situation no policy has yet been written for.

Humans are not replaced. They are repositioned to the boundary, where judgment, presence, and stakes matter more than throughput.