An approach to building business information systems in which the whole enterprise is described as a network of objects with a phased lifecycle, and the participation of humans and AI agents rests on one and the same mechanism.
Business information systems are shifting from one paradigm to another. The traditional one treats the system as a tool that supports a human: it stores data and computes, while the human makes every decision. The new one treats the system as an autonomous executor: it drives objects through their lifecycle on its own, assigns work to people and agents, and makes routine decisions. Human involvement shrinks as the system accumulates data and improves its accuracy.
For systems in the new paradigm you need an information model in which automation is the default and human involvement is a deliberate choice where it is genuinely needed. Phase offers such a model: treat the enterprise from the outset as an automatic cybernetic system, and its objects as entities with behaviour, a lifecycle and relationships — equally legible to an analyst, a developer and an AI agent.
All application logic is described by three and only three primitives. This is the complete vocabulary with which an engineer programs the activity of the enterprise. Everything else — scheduler, adapters, event bus — is engine infrastructure, identical for every object.
An object exists in one of a finite set of phases. It leaves a phase when the criterion — a condition over fields — is met. The transition is instantaneous: condition satisfied → a jump into the new phase.
An object spawns child objects — automatically on a transition or by a participant's decision. Each child is self-contained and linked to its parent by subscriptions in both directions.
An object reacts to another changing phase: when the source leaves a phase, the engine writes a signal into a system field of the subscriber. This is how the enterprise's reaction propagates.
Complex patterns — parallel approval, compensation, branching, timers, hierarchy — are assembled from combinations of the three primitives. No parallel gateways, no "human step" as a separate construct, no compensation events. Fewer concepts means fewer places where the system behaves in a special way.
An employee and an AI agent change an object's phase through one mechanism — writing to a field. There is no special "human step" and "agent step". A field's type sets only the write right; the transition logic is blind to types.
The degree of human involvement is configuration, not architecture. Moving from manual control to autonomy is a change to the object's configuration, not a code rewrite. A team grows autonomy as its trust in the agent grows.
A field holds not just a value but its freshness — when the knowledge was current. The digital model is the best available knowledge about the real object, not absolute truth. Corrections do not rewrite history.
A short vocabulary and declarative configurations with a machine-checkable schema. There is one correct way to do each thing — the agent reproduces the pattern without improvising, and the validator rejects an invalid configuration before it is applied.
Context → decision → outcome → human correction are recorded as a by-product of normal work. A labelled dataset for improving agents appears automatically, with no separate labelling effort.
The transition table is a mathematical object. Before launch one can prove that every phase is reachable, that there are no trap phases, and that safety invariants hold. Autonomy is expanded on proof, not on trust.
Fields are classified by their source of change — an architectural decision about responsibility for data, not a technical detail.
| Field type | Who writes | Purpose |
|---|---|---|
| System | Events: integrations, scheduler, transitions of child objects | Signals from the real world, timers, synchronisation |
| Agent | AI agent via tools | Results of analysis, classification, generation |
| User | Employee via the interface | Decisions, confirmations, data entry |
The engine checks a transition condition without distinguishing whose write satisfied it. A write by an employee, an agent or an external system acts on the object's lifecycle identically — which is why a decision can be handed from human to agent without reworking the model.
Full management of an enterprise has two loops. Action — driving objects toward goals within a given structure. Adaptation — changing the structure itself when it stops achieving the goals. Phase builds both loops with the same three primitives: a goal is an object, a configuration change is an object, and even a subordinate enterprise is a digital twin inside another enterprise.
There is no privileged layer in the model that rebuilds it from the outside. Programming the system becomes ordinary work over an object — symmetric for the engineer and the agent.
The boundary of autonomy is fundamental, not temporary: the agent executes accumulated knowledge and acts from the past, while the new — a new goal, a non-standard decision — is brought in by the human. The management loop is closed and runs on its own, deliberately reaching out to the human where it hits the edge of the known. Not a machine instead of people — a machine that uses human wisdom precisely.