Create, version and troubleshoot workflows
Start with an explicit flow
- In workflow mode, select New and add a manual trigger.
- Choose logic, tool or AI nodes from the node menu and connect the execution path.
- Configure inputs and outputs for each node, save and execute manually.
- Inspect execution records before adding a schedule or API trigger.
A manual trigger is suited to debugging. Sync and async APIs have distinct call and response behavior and are not interchangeable.
Let an agent create or modify a flow
- Describe the goal, inputs, processing steps and expected output in a conversation.
- If Add workflow is displayed, create the entry before asking the agent to generate a flow.
- Inspect the workflow preview and result. Use Open workflow to enter the canvas when available.
- For changes, inspect the before/after diff instead of relying on a textual “modified” response.
- Save, then inspect the version and manual execution result.
A workflow tool calls an enabled API workflow; it is not the same as agent-assisted creation or editing.
Debug and manage versions
- Run debugging from the agent conversation or canvas and inspect failed nodes, inputs and errors.
- Use the version entry to inspect changes and compare versions. Before rollback, check the target version and current unsaved work.
- Rolling back a version does not delete execution history or undo side effects in external systems.
Execution records and AI diagnosis
- Open the target workflow's execution records and select a failed run.
- Inspect node status, errors, retries and associated traces.
- When the failed execution offers AI diagnosis, use it to assist analysis; diagnosis is not automatic repair.
- Return to the canvas, fix the flow and manually verify it again.
Logs and diagnosis context can contain credentials, prompts, business data and external endpoints. Redact before sharing. A visible diagnosis entry does not guarantee that a model is available.
Execute safely
External requests, SSH, file and script nodes interact with real systems. Restrict destinations, directories and credentials; bind parameters and never concatenate user input into commands or SQL. Debug with no-side-effect data before enabling unattended triggers.
