Research

Where our engineering is going.

Six directions make up our development programme. This page explains what each one is for and how they fit together. It describes direction, not a list of finished features.

The six directions Reality Branch Protocol

Technology programme

The six directions.

  1. 01

    Agentic enterprise / AI operating system

    Software agents that carry out work across an organisation, coordinated in one place and kept inside permissions people have set.

    It gives the other directions a way to act.

    Direction of development
  2. 02

    Operational model of the organisation

    One shared description of processes, assets, constraints and objectives.

    It is the common ground. Every other direction reads it and writes to it.

    Direction of development
  3. 03

    Persistent and proactive AI

    AI that keeps context from one day to the next, notices when a condition changes and raises it before anyone asks.

    It keeps the model current and tells the agents and the people when something moved.

    Direction of development
  4. 04

    Adaptive interfaces

    Screens that show each person what they need for the decision in front of them, and change with the situation.

    It is where people see the model, approve and correct.

    Direction of development
  5. 05

    World models and simulation

    A representation of how the real system behaves, so a decision can be tried as a scenario first.

    It lets a proposed action be tested against the model before anyone takes it.

    Direction of development
  6. 06

    Physical AI

    Reasoning connected to vehicles, sensors and infrastructure.

    It carries decisions into the physical world and brings observations back.

    Direction of development

How they connect

One loop, six parts.

The operational model sits in the middle. Persistent AI keeps it current. Simulation tests a change against it. Agents carry out what a person approved through an interface. Physical AI brings the result back from the real world, and the loop starts again.

ObservePhysical AI and connected systems report what is happening. Persistent AI notices what changed.
RepresentThe operational model holds the current state of the organisation and its constraints.
TestWorld models and simulation show what a proposed action would lead to.
DecideAn adaptive interface puts the options in front of the person responsible.
ActAgents carry out the approved steps, and the result is observed in turn.

Programme and products

Direction is not availability.

Nothing on this page should be read as a feature you can buy today. What each product does now is stated on its own page, together with what is real data and what is an example. See the products

Research concept

Working name Concept in research

Reality Branch Protocol - RBP

A proposed way of coordinating scenarios between AI systems and simulators.

ObjectiveCoordinating scenarios between AI systems and simulators, so that several systems can reason about the same situation and the same alternatives.
The problem we look atA delimited one: several systems evaluate alternatives that start from one reference situation, and the reference then changes. Which branches still describe a possible future, and which evaluations have to be done again?
A conceptual exampleThe reference is tonight's charging plan with four chargers. Branch A assumes one charger fails. Branch B assumes an early departure, and its evaluation uses the result of A. At 23:10 the reference changes: a second charger is reported faulty. A still starts from a possible situation but its evaluation is out of date. B depends on A, so it is out of date as well. A third branch that only concerned driver rosters is untouched.
The ideaA shared reference situation, branches that state what they change, and dependencies written down between them, so that an update to the reference marks exactly what has to be evaluated again.
Research questionCan explicit branches and dependencies keep the evaluations of several systems consistent after the reference changes, at an acceptable cost in re-evaluation?
Proposed experimentA small simulated depot and two independent evaluators. Reference updates are injected, and we count the evaluations that remain wrong and the evaluations repeated without need, with and without the declared dependencies. The experiment has not been run.
StatusA concept in research, under a working name. There is no published specification, no software kit and no implementation in our products. It is not an adopted standard, and we make no claim about how it compares with other work.

Contact

Working on the same questions?

If you research or operate systems where these directions matter, write to us.

[email protected]