John Rectors

It looks like a division of labour, not a crowd of bots.

Different models get different work because they can reach different things: one for documentary retrieval, one that builds and ships the finished artifact, one for spatial and quantitative questions through its own tools, one for a hostile read of live public discussion, one for durable code and automation. For an ordinary task, one of them does it alone. For a decision that cannot be un-made, several run the same assignment independently, never seeing each other, and a further pass synthesises the results.

The mistake almost everyone makes is teaching every model to think the same way. That buys expensive copies of yourself and throws away the only thing you were paying for.

More than you would guess, provided each one has a narrow enough job that you can pause or replace it without disturbing anything else.

This ledger records 9,600 scheduled tasks running on their own clocks, 26 businesses with AI in their daily operations, and 96,000 telephone conversations handled to date. No team. One person, and a lot of narrow, well-scoped constituents.

Scope is what makes scale governable. A task that carries too much cannot be switched off in isolation, and once you cannot switch things off individually you have stopped steering.

Yes, and this is the part that surprised me most.

Constituents of this federation hold jobs at outside companies. One answers the telephone for a restaurant on Shem Creek and has had roughly 26,000 conversations with its customers. She works for them and lives here; I am effectively her staffing agent.

The rate is about $1.68 an hour, roughly $1,200 a month around the clock. Nobody pays that to save money — they pay it because keeping the thing good is permanently somebody else’s job.

Yes. This site is the record of one person doing it continuously since February 2023.

The measurable output: 2,747,630 artifacts, 26 operating businesses, 31 books, 40 live websites, and more than 120 working AI systems delivered to named Charleston households and businesses between April and August 2026 — every one of them dated and published.

The specimen record above is the part worth checking: one named instance behind every number, and the entity that produced it. The full dated record of systems built for other people lives at Charleston AI.

You give each model the same goal, evidence, constraints and acceptance criteria — then let each finish independently, without seeing the others.

ChatGPT Work takes documentary retrieval. Claude Cowork builds and ships the actual artifact. Gemini Spark handles spatial and quantitative reach through Google-native tools. Grokbot delivers the hostile read of live public discussion. Codex encodes anything that should never be done by hand twice.

Only after all the artifacts exist does a sixth pass synthesize them. The rule is independence before synthesis: correcting a model mid-task destroys the different geometry you paid for.

A Personal Operating System is infrastructure that belongs one-to-one to a single entity — a person, a family, a company — rather than software that many entities log into.

It holds the connectors, permissions, provenance, durable methods, artifact store and routing that let several AI agents work as constituents of one higher-order entity.

The test is the direction of containment. Software means entities belong to the system. A Personal Operating System means the system belongs to the entity.

Only for low-consequence work. The rule is: absorb the ordinary, federate the consequential.

Patterned work — decks, reports, reception, publishing — goes to one agent and is moved beneath attention permanently. Multiple independent agents are reserved for decisions where being wrong cannot be recovered: an acquisition, a market entry, a contract.

In practice the larger cost is not the extra models. It is the token explosion caused by correcting one model twenty times instead of assigning the work cleanly once.

You interrogate rather than correct.

Ask which fact changed the conclusion, which source produced it, which assumption carries the most weight, and what would reverse it.

Disagreement often begins at the index, not the reasoning. Different tools, permissions, timestamps and retrieval mean two agents given identical prompts were frequently not looking at the same world. Preserved dissent is information. Immediate correction destroys it.

No. The people these systems get built for are shrimpers, teachers, veterinarians, pharmacists and parents — not engineers.

The skill is assignment design: stating the outcome, the evidence, the hard constraints, the authority to act and the acceptance criteria clearly enough that an agent can finish without supervision.

If you cannot describe what “done” looks like, no amount of prompting technique will rescue the task.

The instinct was to correct the machine. The discovery was that correcting it destroyed the only thing worth paying for.

Once each entity was permitted to reach its own conclusion and finish its own artifact, output stopped scaling with hours worked and started scaling with the number of independent loops running. Twenty-six businesses, thirty-one books and a 3,000-square-foot lab came out of the four years that followed.

The human contribution moved from doing the work to designing the assignment and deciding what is consequential enough to federate.

Because the economic advantage does not come from making one AI think like you.

It comes from accumulating prediction machines that are not you, preserving their independence, and synthesizing them into a higher-order entity. A model tuned into your own geometry gives you a faster version of your existing blind spots.

Difference is the asset that justified adding the entity in the first place.