Build your first workflow

Use “customer feedback routing” as an example. Python first performs deterministic cleanup, an Agent classifies and recommends an action, and a rule decides whether a person must review it. This expands the minimal Agent from the quickstart into a production-shaped flow.

Start → Prepare data (Process) → Classify and recommend (Agent) → High risk?
                                                          ├─ Yes → Human Review → End
                                                          └─ No ─────────────────→ End

Define the state contract first

Before connecting the canvas, list the fields every step produces and consumes.

Node	Reads	Writes	Publish to global state?
Start	—	feedback	Yes; later nodes need it.
Prepare data (Process)	feedback	normalized_feedback, is_high_risk	Yes; the Agent and branch need them.
Classify and recommend (Agent)	normalized_feedback	category, priority, recommendation	Yes; review and End need them.
Human Review	recommendation	approved	Yes; the branch or End may need it.

On nodes that produce output, enter these keys in Publish to global state → Published output keys, separated by commas. Output is node-private by default; downstream nodes cannot assume they can read it until it is published and they have been granted read access.

Assign responsibilities

1. Process: read input, normalize fields, apply fixed rules, or query internal systems.
2. Agent: make semantic judgments and return fixed fields such as category, priority, and recommendation.
3. If/Else or Switch: branch only on declared state fields.
4. Human Review: give a person decisions that affect customers, funds, permissions, or public content.

Configure the branch and review

1. Set an If/Else condition such as is_high_risk == true || priority == "high".
2. Connect the “yes” output to Human Review, and the “no” output to End.
3. In Review request, set the content key to recommendation and context keys to category, priority, normalized_feedback.
4. Enable Allow editing if the reviewer should be able to rewrite the recommendation.
5. Connect both the approved and rejected outputs. Do not silently end a rejection; route it to more information, a rewritten recommendation, or an explicit terminal node.

Media placeholder · screenshot workflows/01-state-publication.png
Show the Process node’s “Publish to global state” section with multiple output keys entered. Explain that producing output and exposing it to the workflow are separate actions.

Media placeholder · video workflows/02-branch-and-review.mp4 (45–60 seconds)
Add If/Else, enter the condition, configure Human Review, then run to a pause and approve it to resume.

Design notes

• Make each node do one describable job.
• Define node output before writing prompts and code.
• Use readable state field names rather than stuffing whole documents into every node.
• Put side-effecting operations after review and enable human confirmation for Agent tool calls.

Run three examples before moving to evaluation: ordinary feedback, clearly high-risk feedback, and feedback with missing information. They become your first evaluation cases.