Parallel and Segregated

Ron Reynolds · 2026-09-08 · 5 min read

Somewhere in your company there is a copilot. It arrived last year with a demo that impressed everyone — a chat window on the ERP, an assistant that fills in the CRM faster. The pilot went fine. The renewal went through. And the company works today exactly the way it worked before the demo: same screens, same nightly batch, with a talking window bolted to the side.

That outcome was decided at the moment of purchase. A copilot accepts the existing system as the center of gravity and orbits it. The old software stays in charge of what is possible; the assistant helps you operate it a little faster. Play the tape forward ten years, to the version of your business where agents watch cash, place orders, chase suppliers, and file the paperwork — where the day runs itself while people set policy and read the record — and no amount of copilot renewal gets you there from here.

There is a way that does. Don't renovate. Build beside.

Stand up a separate project, parallel to the business and segregated from its systems, where agents are first-class citizens from the first commit — the operators the system is designed for. Run both. Let the new one take weight as it proves it can carry it.

The case for leaving the legacy alone is what it's made of. Legacy software is shaped for human hands. Every screen assumes eyes. Every workflow assumes a person advancing it; the permission model assumes employees in seats; the integration surface assumes a developer with a quarter to spare. An agent can operate all that the way a person would — slowly, through the costume of a browser session — but nothing in there meets the agent halfway, and the retrofit ceiling is low. What actually carries over is smaller than the balance sheet admits: your data, your customers, your brand, your regulatory record. The plumbing everyone works so hard to preserve is the part worth the least.

Every executive who has watched a rebuild knows the objection, and it is a good one. The greenfield fossilizes too. Two years of build, and the new system ships frozen in the shape of the spec written at kickoff — and five years later it is the legacy, and someone is pitching a rebuild of it. That movie is why companies buy copilots. Renovation at least fails cheap.

The objection dies in a system where agents are first class, because that system is shaped by prompts. You tell it what your business is, the way you would brief a manager you trust, and it takes that shape. When the business changes — a new line, a new market, a season that behaves nothing like last season — you say so, and the shape follows. Reshaping is the same act as running it. The rebuild that fossilizes on arrival was frozen by its own construction method: a spec written once, translated by people who then leave. A prompted system has no kickoff. It tracks the present tense of the business. Replace the old with the now.

Segregated is the half of the prescription that gets ignored. The moment the new system starts syncing into the old database, calling the old APIs, deferring to the old workflow states, it has become one more bolt-on with extra steps, and the legacy's gravity pulls every decision back toward the shape you were trying to leave. The new system stands on its own data and its own rules from day one. So does the team, if people are on it at all — fluency in the old system faithfully reproduces the old system, which is exactly the risk. The new project owes the legacy nothing.

Then the handover runs by function, and the metric is silence. Hand the new system one function at a time, starting with one the old system does badly, and watch the old system's activity on that function fall toward zero. A function that has been quiet for a full business cycle has switched itself. Nobody schedules a big-bang cutover. The old system just goes quiet, one function at a time, until the day someone asks whether anyone still uses it.

ComOS ships in this shape. Every business on it runs a world of its own — its own agents, its own composed platforms, its own data, on hardware it owns if it wants that — admitted to one federation where any world can trade with any other. The separate project can be a machine on a table. The operating system installs on it, the database lives on the machine's own disk, and the link to the wider network points outward only; nothing on the internet can call in, and the legacy stack has no path into the new system at all. A business takes shape by composing platforms (retail if you sell goods, bookings if you sell time, both if you're both), and the owner runs it by talking to it. The segregation the prescription demands is the architecture. A company standing up its parallel project doesn't have to build the ground floor — the separate project is already standing, waiting for the prompt that shapes it into yours. The install takes an evening, and the old system doesn't have to know. Further reading Prompts Are the Agentic Language — the language that does the shaping, and the one-line security model underneath it. The Machine on the Table — the parallel system on a single machine you can pick up and carry. Every Node Is Its Own World — what it means that the separate project is a complete federation of its own.