Preview and debug a graph

Trace the workflow from START through its work and decision nodes to END. Remove stale edges or reconnect them to node IDs that still exist. Use Graph nodes and validation reference to check supported node types and field limits.

The preview request rejects an incomplete graph before it allocates a preview ID.

In the execution input area, select ↻ Sample to reload the schema sample, or select Raw and enter a JSON object. For example:

{
  "ticket": "A customer cannot access the example workspace",
  "priority": "normal"
}

The top-level value must be an object, not an array, string, number, or null. Select Expand JSON when you need the full-screen editor. Agent Studio checks the object against the selected skill version before starting the preview.

Select Ghost Preview. Agent Studio validates the graph, confirms that the skill version exists and belongs to a skill you can access, validates the input, and registers a temporary preview.

On success, a notification reports Ghost preview started and explains that the graph is dry-running without persistence. The canvas opens the trace against the returned preview-… ID.

Keep the trace connected while the workflow runs. The first SSE subscriber starts the dry-run, and the stream then reports node starts and ends, traversed edges, execution status, router decisions, loop iterations, and other node-specific details produced by the workflow.

Use the sequence to find the first divergence: a node that fails, a router that chooses the wrong label, a loop that reaches an unexpected iteration, or an edge that never receives a traversal event. The stream sends a heartbeat every 15 seconds to keep the connection active.

Wait for an execution:status event. The stream closes when the preview reaches COMPLETED, FAILED, CANCELLED, STEP_LIMIT_EXCEEDED, or PAUSED_FOR_APPROVAL.

COMPLETED confirms that this input reached a successful terminal status. The other statuses identify the class of interruption; inspect the immediately preceding node and edge events for the point of failure or pause.

Select BACK TO EDIT after you have recorded the failing node, route, and status. Correct the graph, prompt, tool configuration, or input, then start a new Ghost Preview.

A new preview is the reliable way to retest. After a preview finishes, its temporary session is deleted, so reconnecting does not replay missed events.