Test the agent
By the end of this page you will know how to run test executions, inspect node-by-node results, and debug failures in your Task Agent.
Prerequisites
- A Task Agent with at least one node between Start and End (Create a Task Agent).
- Admin or Developer role.
- The agent must be saved (unsaved changes cannot be tested).
Opening the Test panel
Click the Test button
In the Agent Builder toolbar (top-right), click the Test button. A full-screen Test Execution panel opens, overlaying the canvas.
The panel has four sections:
- Header — Shows “Test Execution” with status (Ready / Running / Completed / Failed), execution time, and node progress (e.g., “0 / 11 nodes”).
- Input sidebar (left) — Lists all input variables defined in the Start node with their types, plus the Start Test button.
- Execution canvas (center) — A read-only copy of your graph showing real-time execution status per node, with a Running / Completed / Failed legend.
- Agent Console (bottom) — Streams execution logs in real time once the test runs.
Fill in input variables
In the left sidebar under Input Variables, fill in values for each variable. Required variables are marked with an asterisk (*). Variable types (string, number, boolean) are shown in parentheses.
Use realistic test data that exercises all branches in your workflow. Test edge cases (empty strings, large numbers) to verify error handling.
Start the test
Click the Start Test button at the bottom of the input sidebar. The panel status changes to “Running” and the execution timer starts.
Watching execution in real time
As the agent runs, each node on the execution canvas updates its visual state:
| State | Indicator | Meaning |
|---|---|---|
| Running | Animated border | Node is currently executing |
| Completed | Green checkmark | Node finished successfully |
| Failed | Red X mark | Node encountered an error |
The header updates the node progress counter (e.g., “7 / 11 nodes”) and the elapsed time in real time. Task Agents often complete in well under a second, so the running animation is brief — the canvas then settles into the final per-node state (green for completed, red for failed), and the Agent Console shows the full log.
Inspecting results
View node output
After execution completes, click any node on the execution canvas to inspect its result. A detail panel shows:
- Input — The data received by the node.
- Output — The data produced by the node.
- Duration — How long the node took to execute.
- Error (if failed) — The error message and stack trace.
Check the final output
Click the End node to see the agent’s final output — this is what would be returned to the caller in production.
Review the Agent Console
The Agent Console at the bottom of the test panel shows a chronological log of execution events with timestamps and severity tags (runtime, oote, sys, stream) — including the resolved input payload, per-node updates, and a final “Agent execution completed” line with the total duration.
Debugging failures
When a node shows a failure (red X):
- Click the failed node to see the error message.
- Check whether the input data was correct — a common cause is a missing or null variable from an upstream node.
- For Code nodes, review the console output and stack trace.
- For Tools Call nodes, verify the tool configuration and endpoint availability.
- Fix the issue in the main canvas, save, and re-run the test.
The Test panel uses the saved version of your graph. If you make changes, click Save before re-running the test.
Closing the Test panel
Click the X button in the top-right corner of the Test Execution header to return to the main canvas view.
Common issues
- “Start Test” button disabled — You have unsaved changes. Save the agent first.
- Test hangs on a node — The node may be waiting for an external service (HTTP call, tool execution). Check timeouts and service availability.
- “Variable not defined” error — A node references a variable that isn’t produced by any upstream node. Check your variable wiring.
- Inconsistent results between runs — If your agent calls external services or uses non-deterministic LLM outputs, results may vary. Use fixed test data where possible.