generated: '2026-07-19'
method: searched
source: https://flow-ai.com/docs/concepts
summary: >-
  Cross-cutting runtime semantics of the flowai-harness data-agent runtime. Flow AI
  is a Python/Rust SDK and embedded runtime, not a hosted REST API, so these are
  runtime-contract conventions rather than HTTP request/response conventions.
authentication:
  style: bring-your-own model-provider keys (env/config); Production Studio adds SSO/RBAC.
  detail: authentication/flow-ai-authentication.yml
tenancy:
  model: TenantIdentity via define_tenant(...); runtime isolation keyed by resource_id.
  purpose: scopes catalog, storage, and runs to a customer/workspace/environment.
plans_and_actions:
  model: typed plan and action schemas validated by the runtime and executed as auditable
    state machines; plans are persisted and status-tracked.
approvals:
  model: runtime approval gates pause sensitive work until the host application approves
    or rejects an action before the action dispatcher applies it.
references_and_glimpses:
  model: references are typed handles to stored values; a glimpse is the small summary
    stored beside a reference so agents can reason about a payload without resolving
    the full value. Used to pass large/sensitive/intermediate data between tools, plans,
    executors, and host code.
streaming:
  model: runtime events are an async-iterable observable stream (text, reasoning, tool
    calls, plan/action status) rendered by an application or Studio.
prompts:
  model: layered_prompt(...) with deterministic rendering; explicit layers including
    domain_knowledge and operational_rules.
mcp:
  model: runtime tools can be exposed as MCP servers over stdio or Streamable HTTP.
  detail: mcp/flow-ai-mcp.yml
idempotency:
  supported: false
  notes: No documented HTTP idempotency-key contract (not a hosted REST API).
cross_links:
  authentication: authentication/flow-ai-authentication.yml
  changelog: changelog/flow-ai-changelog.yml
  sandbox: sandbox/flow-ai-sandbox.yml