Debugging an Agent¶
Dataiku provides several ways to investigate errors and unexpected agent behavior. You can review agent logs, inspect traces, and store interactions for later analysis. When testing a visual agent, you can also use live debugging to follow its execution.
Agent logs¶
Agent logs help you troubleshoot errors and understand what happens during execution.
Open the Logs tab of your agent to view its logs. These logs are updated each time the agent is called through the LLM Mesh.
For Code Agents in particular, you can add Python print() statements to your code to record intermediate values and follow execution. Their output appears in the agent’s logs.
For more detailed debugging, you can also use:
Tracing to inspect nested processing steps, inputs, outputs, and metadata. Code Agents can add custom spans to their traces.
Interaction logging to store agent interactions in a dataset for later review, debugging, and evaluation.
Live debugging¶
Live debugging shows the execution steps that produce a visual agent’s response during a single run.
It is available from the Design tab of saved Simple Visual Agents and Structured Visual Agents, where you can inspect a run while testing an agent.
Use it to:
Follow the execution of a run while it is still in progress
Identify which tools, blocks, and sub-agents were used
Inspect the inputs, outputs, and metadata of block and tool executions
Review changes to the agent state and scratchpad
Troubleshoot unexpected agent behavior
Enable live debugging¶
When using the test chat or sending a test query, enable the Debugger checkbox to display execution events. Select an event to view more details in the debugger panel.
Limitations¶
Live debugging is available only in the Design tab of saved agents.
Live debugging is not available for Code Agents.
Execution reporting for Parallel blocks is not live.