This walkthrough explains stable platform ideas. Business entity and metric meanings belong to the detailed domain model, runtime and backend choices to the detailed architecture reference, exact routes and parameters to the full API reference, and current evidence and gates to engineering status.
AgentFlow treats fresh operational context as the default. Batch jobs can still exist, but the agent-facing model is built around bounded or unbounded event streams that flow through the same validation and enrichment concepts.
This matters for agents because stale state can cause the wrong action: answering a support ticket from an old order status, escalating an already resolved incident, or recommending inventory that has just changed.
Agents use stable business entities, metrics, constrained questions, search, lineage, and contracts instead of calling storage tables directly. The detailed domain model owns their names and business interpretation. The semantic layer maps those concepts to backing storage so client code stays small while the runtime can evolve behind the API.
AgentFlow keeps schema contracts explicit so callers can reason about response shape and compatibility. The full API reference owns the exact listing, version, diff, and candidate-validation routes.
Contracts are also a boundary against accidental data drift: a new field can be additive, while a removed or type-changed field needs a migration path.
The local path optimizes developer feedback; a production-shaped path exercises external processing, storage, deployment, and observability boundaries. Both preserve the same validation and semantic contracts. Architecture owns the component choices; the deployment walkthrough selects a runnable path.
Production operation still requires owner-owned hostnames, secrets, cloud account wiring, retention policy, monitoring, and rollback decisions.
Natural-language query execution is constrained. The server validates translated SQL and restricts the operation to allowed read-only shapes before executing it. Entity and metric lookup paths use parameterized queries for untrusted values.
Checked-in gates and local evidence establish only the claims recorded in engineering status. External infrastructure, security assessment, retention, and compliance controls require separate owner evidence.