The Observability tab in Zoho Zia Agent Studio provides a comprehensive view of how your deployed agents are performing. It operates across three integrated layers: a dashboard of headline metrics, a session list for tracking individual invocations, and a detailed execution timeline for diagnosing specific runs.

Performance Metrics Dashboard

The dashboard displays eight key performance indicators, refreshed every 30 minutes, giving you a consistent view of how the agent is operating over time:

• Session and execution counts. The total number of invocations and execution runs during the monitored period.
• Performance latency. Average execution time alongside the p95 percentile measurement, which helps identify outlier runs that are taking significantly longer than typical.
• Success rates. Separate success metrics for tool completion, overall execution, and LLM call reliability, giving a granular view of where failures may be occurring.
• Token consumption. Combined input and output token usage across all executions in the period, useful for managing costs and understanding workload.

Session Monitoring

Individual agent invocations appear as rows in the session list. Each row shows the session and version identifiers, creation timestamp, most recent activity time, token processing totals, and step counts broken down by execution, LLM, and tool categories. Selecting any session opens its detailed metadata and full execution history.

Sessions initiated by triggers are marked with "Triggers" as the Execution Source, making it straightforward to separate automatic runs from user-initiated or API-driven invocations.

Execution Analysis

The Step Timeline is the core tool for understanding exactly what the agent did during any given run. It presents the agent's actions in sequence and includes the following step types:

• Initial request. The exact prompt or trigger input the agent received.
• Prompt assembly. How the agent combined its instructions, context, and any gathered data before passing the request to the language model.
• LLM processing. The model call itself, including the response generated.
• Tool and API calls. Full request JSON and response outputs for each tool the agent invoked during the run.
• Final response. The answer delivered to the user, alongside the agent's step-by-step reasoning explanation.

Performance Diagnostics

Use the Observability data systematically to identify and address issues:

• Monitor success rate metrics across tools, executions, and LLM steps for any decline from baseline.
• When tool success rates drop, open individual session timelines to examine the request JSON and response outputs for the failing tool calls.
• Review step durations in the timeline to identify which steps are contributing most to latency.
• When accuracy problems arise, read the reasoning sections in final response steps to understand how the agent arrived at its answer and where its logic may have diverged.