Use cases connect runtime layers to concrete tasks, risks, contracts, failure modes, evidence, and success metrics. They are reference patterns, not claims about a shipped ARuntime product.
Key takeaways
- Add runtime controls in proportion to duration, state, side effects, data sensitivity, and consequence.
- Keep the domain system of record outside model memory.
- Use deterministic software when the workflow is stable and completely specified.
Use-case explorer
[ar_use_cases]
When runtime control is justified
Runtime controls become valuable when work spans several steps, crosses systems or tenants, invokes tools, changes state, pauses for people, needs recovery, or must be reviewed. The architecture should identify the minimal controls needed rather than importing every agent feature.
When to use a simpler system
| Situation | Prefer |
|---|---|
| One read-only generation with no sensitive data or durable state | Direct model API with validation and basic telemetry |
| Fully specified transformation | Deterministic code or codemod |
| Known business rules and forms | Traditional application workflow |
| Search and retrieval without synthesis | Search UI and source links |
| Fixed scheduled integration | ETL/job scheduler |