This comparison distinguishes runtime categories by execution unit, input, output, state, policy, tools, traces, and typical boundary. It compares responsibilities, not universal product quality.
Category comparison
| Category | Primary unit | Typical input | Typical output | Owns tools | Owns durable state | Enforces application policy |
|---|---|---|---|---|---|---|
| Compiler/graph runtime | Graph/module/dispatch | Model graph and target | Executable/tensors | No | Artifacts only | No |
| Inference engine | Model invocation/sequence | Model and tokens/tensors | Tokens/tensors | No | KV only, usually ephemeral | No |
| Model server | Network request | API request and model version | Network response | No | Deployment state | Service-level only |
| AI gateway | Proxied request | Model/tool traffic | Routed request/decision | May mediate | Limited | Edge/gateway policy |
| Workflow engine | Workflow step | Events and activities | Durable state/result | Via activities | Yes | Generic workflow policy |
| Agent framework | Turn/node/handoff | Messages, agents, tools | Messages/tool requests | Declares | Varies | Varies |
| Agentic application runtime | Task/tool/approval | Actor, task, context, authority | Controlled work/evidence | Yes | Yes | Yes |
| Product application | Domain transaction | User/business intent | Domain outcome | Through runtime | System of record | Owns business rules |
Boundary notes
“Owns” means the component provides the lifecycle and enforcement, not merely an integration. A model server can call a policy service without owning application policy. An agent framework can store checkpoints without establishing tenant isolation or compensation.
Products spanning layers
[ar_diagram id=”cross-layer-product-example”]
Record a product’s primary category, secondary categories, and covered layers. Compare only the subset relevant to a shared scope.
Some apparent product comparisons are category errors. GGUF vs GPT demonstrates the difference between an artifact format and a model architecture.
How to compare
Define workload, versions, hardware, deployment, data boundary, SLO, features, exclusions, and evidence. Use product comparison scope before making a table or recommendation.