Agent steps

See what your agent did during each conversation turn in real time.

Agent Steps is a panel in the dev console Chat view that shows you exactly what happened during each conversation turn. While you chat with your agent on the left, the right panel shows a live breakdown of every action the agent took to produce its response.

What it shows

For each turn, Agent Steps displays:

• User message that triggered the turn
• Tool calls with input and output for each tool the model called
• Knowledge searches showing which queries ran and what passages matched
• LLM reasoning including the code the model generated
• Messages sent back to the user
• Errors if anything failed during execution
• Duration for each step and the total turn

Each item is expandable. Click on a tool call to see its full input and output. Click on a knowledge search to see the matched passages and their scores.

When to use it

Agent Steps is the fastest way to debug during development. Use it when:

• The agent gives a wrong answer and you want to see which tool it called (or didn’t call)
• You want to verify the agent is using your knowledge base correctly
• A tool call fails and you need to see the error
• You want to understand the model’s reasoning path
• You’re tuning instructions and want to see how changes affect behavior

Traces link

Each turn in Agent Steps links to its full trace in the Traces view. Click the trace link to see the complete span timeline with detailed telemetry. See Debug with logs and traces for more on the Traces view.