AI Workflow Designer Quality Review
The AI-assisted workflow designer creates a draft from your goal and the resources available in your workspace. The Review step now explains whether that draft is structurally ready, what still needs attention, and what its stages would look like before anything runs.
The review and dry run are static. They do not execute agents, call providers, read files, activate the workflow, or spend workflow credits.
Read the quality score
The score is a deterministic second opinion, not a promise that the AI’s business reasoning is correct:
| Band | Score | Meaning |
|---|---|---|
| Excellent | 90–100 | No significant structural findings were detected. |
| Good | 75–89 | The draft is a useful starting point, but review the findings. |
| Needs review | 55–74 | Design, resource mapping, or review coverage still needs work. |
| Blocked | 0–54 | Repair or regenerate the graph before treating it as a viable design. |
The normal workflow validator remains authoritative for saving, activation readiness, and required configuration. A high score does not bypass workspace permissions, resource readiness, approval routing, or cost limits.
Understand findings
Findings explain the issue in plain language and include a suggested correction. Common checks include:
- a reachable entry path, terminal step, and finite graph;
- duplicate IDs, broken connections, and unreachable steps;
- instructions and expected outputs for executable steps;
- assignment to an approved local agent, group, or role;
- synthesis or review coverage after parallel work;
- a human approval step for sensitive or consequential work;
- notification channel and message configuration;
- failure connections and behaviors that the current runtime does not enforce as recovery branches.
Select Improve draft to send the findings through the existing revision flow. The original goal and the draft-only status are preserved; you can inspect the new result before accepting it.
Use the static dry run
The dry run lays out the finite, topological path as stages. Each stage is labelled as sequence, parallel, condition, review, or terminal. It also shows the number of parallel and conditional branches and the maximum depth.
Use it to answer questions such as:
- Where does this workflow start?
- Do parallel workers converge before delivery?
- Where is review or human approval expected?
- Does the graph have a clear terminal result?
If the preview says that no finite path can be simulated, inspect cycle, entry, terminal, and unreachable-node findings before accepting the draft.
What you still need to do
After reviewing the draft:
- Map any remaining steps to approved workspace resources.
- Configure readiness requirements, notification destinations, and approval routing.
- Review the run-cost estimate and workspace policy limits.
- Save the draft, then activate it only after the normal validator reports that it is activation-ready.
The quality layer does not prove factual correctness, test external integrations, or simulate provider responses. Treat it as an explainable design review that makes the next human decision easier.
See Agent Orchestration for execution, review, and runtime governance.