How AI Flow works — end-to-end
This is the daily loop in the UI.
Inner loop (build and test)
Edit graph / prompt / tools → bind a Dynamics model if the flow needs an LLM → configure retrieval or APIs → Playground → inspect the answer (and View traces) → change the graph → Save draft → rerun.
A. Get in
Sign in → set AI Project → marketplace or left menu → AI Flow. Check the flow list so you know what already exists.
B. Create or open
- + New Flow → choose Flow (visual) or Native (direct LLM call) → Continue.
- Or click an existing name such as
network-flow-agent.
C. Build on Graph
Drag components from the left Components list. Connect output ports to input ports. Fill required fields (red asterisk). Use the node play / pencil / trash icons to run, edit, or delete a step.
D. Ask
Playground → type a message → Send. Keep questions aligned with what the flow is built for (for example network KPI JSON into a notification agent).
E. Release
Save draft while iterating. Publish when the graph should be snapshotted. Check Versions and Settings after that.
Words you will see
| Word | Meaning for you |
|---|---|
| Flow | Visual graph of components (the usual kind) |
| Native | Direct Dynamics LLM / service call, no canvas |
| Draft | Unsaved-as-version working copy |
| Publish | Immutable snapshot you can roll back to |
| Playground | Chat test harness for the current graph |
| Components | Building blocks (models, parsers, APIs, agents) |