After a task agent step runs, you can review a record of the agent's reasoning, tool calls, and final answer. Use this information to verify that the agent called the correct tools and to identify failed tool calls.
Before you begin
- A workflow with at least one task agent step must have completed a run.
Procedure
- Open the execution detail view for a completed workflow run on the workflow builder canvas or from the Workflow Runs page.
- Click the Task Agent activity in the activity list below the canvas, or click the task agent step on the canvas.
The step details panel opens. Because this is a task agent step, two tabs appear above the output area: Input/Output and Agent steps.
- Click the Agent steps tab.
A header bar displays summary statistics for the agent's execution:
- Model: The large language model (LLM) that processed the request.
- Tokens: The total number of tokens consumed across all reasoning and tool call steps.
- Trace time: The total wall-clock duration of the agent's execution.
- Tool calls: The number of tool calls the agent made during execution.
- Scroll through the agent steps to review the execution sequence.
Steps appear in chronological order and include three types:
- Reasoning blocks: The agent's chain-of-thought text, showing how it interpreted the prompt and decided what to do next. Each block displays its token count and duration as a combined string (example: "100 tokens · 400ms").
- Tool call cards: Each tool the agent invoked during execution. Each card shows the tool name, the request the agent sent, and the response it received. Click Show input to expand the full JSON input the agent sent to the tool. Token count and duration render as a combined string on the right side of the card header.
- Final answer: The agent's concluding response after all reasoning and tool calls are complete.
- To identify failed tool calls, look for tool call cards with a red border and a Failed label next to the tool name.
A failed tool call means the tool returned an error or did not respond. The agent may have continued with other tools or provided a response based on the information it had.