Architecture

Assembly is designed as a coordination layer that remains independent from specific AI models, agents, or hardware platforms.

Its architecture focuses on separation of concerns, allowing intent, planning, and execution to evolve independently while remaining tightly coordinated.

High-level components

At a high level, Assembly is composed of:

• an Intent Manager, responsible for storing and updating system intent

• a Planning Engine, generating and revising executable plans

• an Actuation Interface, dispatching actions to agents or machines

• a Feedback Channel, collecting execution outcomes and state updates

These components communicate through events rather than synchronous calls, enabling loose coupling and resilience to failure.

State and events

Assembly maintains minimal shared state. Intent and execution state are persistent, while plans are treated as ephemeral artifacts.

State changes are propagated through events, allowing multiple components to react independently without centralized control flow.

This event-driven design enables:

• concurrent planning and execution

• partial system recovery

• isolation of failures

Stateless coordination, stateful execution

Where possible, coordination logic remains stateless. This reduces complexity and allows the system to scale horizontally.

Execution and environment state are treated as stateful and external. Assembly observes and reacts to them without assuming full control.

System boundaries

Assembly does not perform perception, low-level control, or learning. It integrates with existing systems through well-defined interfaces.

By defining clear boundaries, Assembly remains adaptable and avoids coupling coordination logic to specific implementations.