Platform

Something worth writing to.

The mind’s operating surface, and its feed. An avatar is only as capable as what it can take hold of, and a mind learns only what reaches it. Every part of this exists to be authored, and every part of it reports back.

Work with us Who writes to it

A runtime runs programs. Ours runs the operation.

The program is the operation itself, and agents write it. It executes durably in the cloud, and the Companion is the same runtime standing where the machine is. What we do not do is reach inside a body: the intelligence and the control loop in a humanoid or a vehicle stay its own. It arrives already able to think, and what it needs from us is a world to think about — an identity, a role, work, claims, delegated authority and a memory. We build the operation. The body brings the mind. And this is the platform’s other job, the one that makes it more than a runtime: it is the operating mind’s feed. It learns from a desensitised, abstracted model of the work: the shapes and first principles of how operations flow, never a site’s raw record. Your operational record stays in your organisation. The next operation it runs starts from what the last one taught.

Programs

Scenarios. Plain TypeScript carrying one operation's world, records and logic, written by agents and people alike.

Processes

Missions. Each admitted once, journaled, and alive across a restart of the thing that runs it.

Memory

The world model. Units, spaces, claims and custody, one typed picture everything reads and writes.

System calls

Verbs. The only way a program touches a machine, a door or a person, and each one typed.

Six things the platform knows about.

A primitive earns its place by carrying unchanged into an arena we have never seen. Nothing here names a kind of machine, which is why the same core runs a humanoid on a line, a drone survey over a district and a desk of digital workers. It is also why a model is a first-class actor here rather than a spectator.

Mission

A unit of work with a lifecycle, admitted once per key and executed on a durable journal. It outlives the process that started it.

Claim

A hold on a resource over an interval. Reservations are what the platform granted; custody is what the world reported, and the world is late, contradictory and sometimes silent.

Unit

The thing that acts and the thing acted upon are one primitive, so a carrier carrying a carrier needs no new type. An AMR, a drone, an arm, a conveyor, a person, an agent.

Verb

What a scenario may invoke. Move means drive there and arrive. A verb that returns on dispatch is a verb that lies about the world.

Space

Frames, poses and symbolic locations with the transforms between them, from a bay on a line to a district on a map. A dock keeps its name after the site is re-surveyed.

Scenario

Data that carries code. Pinned by the digest of its bundle, promoted by moving a binding, rolled back in one call.

The world model of the Aurora Care Campus operation: a type browser on the left with two additions the mind proposed from the record, a graph of typed objects, and a trolley whose custody is contested with the two reports that disagree — The ontology, with two additions the mind proposed from watching the operation: a property units keep reporting and a link the planner uses that the schema never had. Both are drafted against v8 and neither is published, with the Publish v8 control sitting under them unpressed.

Whatever it is, it answers the same verbs.

An agent cannot run an operation it has to learn six vendor APIs to touch. Making everything behave as one system is not the headline. It is the precondition for the headline.

Humanoids and mobile robots

General-purpose bodies and every wheeled, tracked or legged unit, each behind the same contract.

Autonomous vehicles and drones

Units whose work spans a district rather than a site, under the same missions and the same claims.

Infrastructure

Lifts, doors, barriers, chargers and signals as first-class resources. A call that takes four messages on the wire takes one verb in your scenario.

Industrial AGVs

VDA5050 spoken natively, so a compliant fleet joins without a translation layer of your own.

Open middleware and existing fleet managers

Whatever already runs your operation becomes one more capability rather than something to rip out.

Your own hardware

One adapter against a typed port. Vendor N+1 changes zero platform lines, and the diff proves it.

Reservations on one contended gate of the Riverside Cold Chain operation, drawn over an interval: one unit admitted, one queued behind it, one deferred; the decision record naming the term that settled each; a scope lease with its expiry and reaper; and two totes whose custody reads UNKNOWN — Reservations on the one contended gate, solved as a cohort: agv-23 admitted, agv-24 queued behind it inside its own priority, agv-26 deferred under the charge floor. Refusing agv-23 would re-solve the whole set rather than free the slot, which is why one move is never dropped from it alone. On the right two totes report nothing at all, so custody reads UNKNOWN — written down rather than inferred away, and neither tote is moved by anyone until one report settles it.

The capability plane of one operation: twelve units — a shuttle, delivery units, survey drones, a lift bank, a door, a barrier and a person — each with its adapter, its verbs, the delivery guarantee and the transfer model it declares; a catalogue of the kinds the platform actuates; and a command ledger where a vendor re-sent one task and it was merged — A lift, a door, a barrier, a drone, a shuttle and a person, each behind a typed contract none of us wrote, all answering out of one verb surface. A capability re-sent the same task 63 seconds after the first report, past the hatch that ignores anything under a minute; it was matched on a key the adapter had to synthesise because that endpoint issues none, nothing was re-issued, and the unit drove once — the ledger block, top right.

Point it at work it has never seen.

A humanoid joins the way a forklift does, and a vehicle crossing a district is the same primitive as an arm on a line, because nothing in the core knows what any of them is. A scenario is data an agent writes, so the next operation this runs is one nobody designed it for, and the platform will not be able to tell the difference.

Humanoids and embodied AI

General-purpose bodies walking into work shaped for people. The model is ready long before the operation around it is. That gap is the whole problem.

Cities that run themselves

Autonomous vehicles, drones and public infrastructure as one system across a district. Not a fleet at a time, not a site at a time.

Warehouse and industrial

The most automation already installed, the least coordination between the pieces of it.

The AI workforce

Digital workers beside physical ones, under one model, one set of permissions and one record.

Four pictures, because a page needs pictures. The core has no idea any of them exist.

Plain TypeScript. Live. Versioned by content.

A scenario is the ontology, the records and the logic for one operation, in a language your team reads and every model writes. Its identity is the digest of its bundle, never a tag, because a tag can be moved. A mission in flight stays on the version it started on.

Where it runs at the edge

A scenario the Surveyor agent wrote for the Riverside Cold Chain operation: the gate policy in TypeScript with the agent's own line on why this shape, its diff against the bound version, the lockfile, the specs including one written after rehearsal, and the binding a person may promote — An agent wrote v19, said why this shape rather than the obvious one, and wrote a fourth spec after the rehearsal found what the author had not. It cannot promote its own binding, and the row that refuses it names the envelope; a person moves it.