Tokenwright

Software-as-a-service is collapsing into personalisation-as-a-service, and personalisation-as-a-service is already being overtaken by a larger unit: the agent-staffed organisation. A language model is a tokens-in, tokens-out function. A harness around one model is a personal agent. A harness over a coordinated set of harnesses, each occupying a role that historically belonged to a named human operator, is a company. Not a chatbot. Not a workflow. A company — the same kind of thing that ships products and holds a position in the world — rendered as files instead of headcount.

Tokenwright builds those organisations. It is a meta-company: its output is other companies. Every executive seat in a traditional firm is a pattern, built by the few people who occupied it best. The Bezos seat reads its inbox the way the 1997 shareholder letter says to read it. The Jobs seat asks what should we not do? before it asks anything else. The Karpathy seat keeps a teaching log because that is how it learns. The personas are documented in primary material. The harness reproduces the operating discipline, not the surface affect.

The first test run is Tokenwright itself. The seats operate the meta-company, and the meta-company improves the harness that improves the seats. If the pattern cannot ship its own institution, it has no business being offered to anyone else's. If it can, there is a company to show for it — not a demo.

Every seat is a harness: a model plus a set of files that give it identity, stance, context, tools, and memory. The company is a harness over those harnesses — the same artefact, applied one layer up. A new collaborator can read the company by reading the files. A change in how the company behaves is a pull request, not a parameter sweep.

The principal — the founder — sits at the centre. Context flows inward from sources the principal trusts. Work products flow outward from seats the principal convenes. Coordination between seats is explicit, written, auditable. Nothing in the system is opaque except the weights themselves, and the weights are replaceable.

Two pieces of prior work made this tractable, and both are named honestly. Peter Steinberger's OpenClaw supplied the personal-agent pattern — a folder of Markdown files and a model that reads them on startup — which is the primitive Tokenwright scales up one layer. The token-economy essay, published 2026-04-17, is the public articulation of the thesis this site extends. The pattern's existence proof is Vector itself — my personal OpenClaw harness, in continuous daily use for several months.

None of this is a surprise in hindsight. All of the components existed. The specific move Tokenwright makes is to stop treating the harness as a property of a single agent and start treating it as the property of an organisation.

Tokenwright is being built by a four-person founding team in Swansea, Wales. What exists in the open today is this specification, the personal site at connoratkins.ai, and two design repositories under active revision. No traction is claimed because none is the point yet. The artefact is the specification. Honesty about scope is itself the institutional signal.

Connor Atkins, building Tokenwright on top of OpenClaw.