This diagram presents the seven-layer AI runtime stack as an accessible reference. The visual is accompanied by structured text and can be downloaded as SVG for architecture reviews.
Key takeaways
- Lower layers execute model programs; middle layers serve and distribute them; upper layers govern tasks and domain outcomes.
- Products can span layers.
- Identity, policy, evidence, versioning, cost, and recovery cross the stack.
Seven-layer diagram
[ar_diagram id=”seven-layer-stack”]
How to read it
- Start with the unit of work you need to control.
- Trace downward to the execution resources it requires.
- Trace upward to the product or human outcome it serves.
- Mark every state boundary, network boundary, tenant boundary, and irreversible side effect.
- Assign an owner for detection and recovery at each failure boundary.
Execution paths
Model execution
[ar_diagram id=”model-execution-path”]
Request execution
[ar_diagram id=”request-execution-path”]
Download and reuse
The SVG preserves text, title, description, and high-contrast semantics for light, dark, and print contexts.
[ar_downloads file=”seven-layer-stack.svg”]
Use the stack in an architecture review
Begin with the product outcome and trace downward rather than beginning with a favored framework. Name the unit of work at each layer: graph, kernel dispatch, inference request, token sequence, distributed task, tool action, or business transaction. Then record the state that crosses the boundary, its owner, its lifetime, and the contract used to validate it.
A useful review marks four kinds of boundaries directly on the diagram: process or host boundaries, tenant boundaries, trust boundaries, and irreversible side-effect boundaries. The same product may occupy several layers, but each responsibility still needs an accountable owner. A platform that both serves models and brokers tools should expose separate operational signals and authorization contracts for those duties.
Boundary questions
- Which layer owns admission, overload, cancellation, and deadlines?
- Where are model artifacts, compiled executables, caches, session state, durable memory, and evidence stored?
- Which interfaces are stable enough to test independently?
- Where can data cross a region, device, tenant, or organizational boundary?
- Which failures can be retried, which require reconciliation, and which must stop?
- What evidence remains after an execution worker or model deployment disappears?
Use the diagram with the textual layer definitions so review decisions remain understandable in print, search, and assistive technology.