Who, how and where.
The agent is who holds the responsibility. The substrate is how the task receives reasoning or execution capability. The node is where the physical capacity lives.
Keeping these layers separate allows a persistent agent to use an eligible general model, specialist model or bounded program without treating each engine as a new organizational identity.
Eligibility is a deliberate decision.
A substrate needs an explicit identity, version, declared capabilities and allowed scope. Qualification should match the task and the environment where it will be used.
- Pinned model or program artifact and version
- Supported task and interface
- Permitted data and operating scope
- Resource requirements and observed readiness
- Evaluation evidence and limitations
Distribution follows the workload.
A substrate may execute locally or as part of a configured cluster environment. Placement must account for availability, data boundaries, memory, interconnect and the selected execution framework.
Multiple nodes do not automatically share one accelerator memory pool. Multiplex and the deployment must agree on the capacity they actually have.
Admission, execution and quality are separate.
Registration makes a substrate known. Admission makes it eligible within a scope. Execution evidence shows that a particular task ran. Evaluation establishes a supported quality claim.
The civilization chart should preserve these distinctions when it connects the logical organization to its physical runtime.