Agent auditing and trajectories

​

Overview

Poolside lets you investigate, debug, and audit agent behavior from the Poolside Console by reviewing usage metrics, session history, and detailed execution trajectories. You typically review this information in sequence: from the agents list to an agent, then to a session, and finally to its trajectory. This progression lets you start with broad signals and focus in on the exact execution details you need. Using this information, you can:

• Investigate why an agent produced a specific result
• Identify failed or error-prone agent runs
• Review which tools the agent used during execution
• Identify when and who approved tool usage
• Trace behavior across sessions with multiple steps or turns
• Debug and troubleshoot individual executions

​

Before you begin

To use the features on this page, you must belong to a team with the tenant-admin role. If you have the View Agent Sessions permission instead, you can view your own agent session trajectories from your IDE using the /trajectory command. For more information, see Commands. For a complete list of permissions, see Permissions reference.

​

View agent usage across the organization

Use the Agents page to review high-level usage across agents and identify activity patterns across your organization.

​

View all agents and usage

1. In the Poolside Console, navigate to Agents > Agents.
2. View the list of configured agents.
3. Review per-agent usage metrics, including sessions, total tokens, and steps.
4. Compare usage across agents to understand activity levels and patterns.
5. Select an agent to view its details and associated sessions.

​

View inference metrics

Enable Inference metrics to overlay time-to-first-token (TTFT) latency data on the usage chart. This is available on the Agents page and on individual agent pages. TTFT can help identify early signs of GPU inference saturation. When inference nodes are under heavy load, queuing and resource contention can delay the start of generation, causing TTFT to increase even when overall throughput appears stable. Use these metrics to determine whether agent usage patterns warrant closer monitoring of your inference deployment.

1. In the Poolside Console, navigate to Agents > Agents.
2. Toggle Inference metrics on.
3. Review the latency lines overlaid on the usage chart:
   • Avg TTFT: Average time to first token
   • P50 TTFT: Median latency
   • P95 TTFT: 95th percentile latency
4. Hover over the chart to see exact TTFT values.

​

Export usage data for all agents

Export usage metrics for all agents to a CSV file for reporting or analysis.

1. In the Poolside Console, navigate to Agents > Agents.
2. Select the date range for the usage data you want to export.
3. In the Metrics table, click Export Data to download a CSV file.

The export includes per-agent session counts, token usage, step counts, and TTFT latency metrics for the selected date range.

​

Review usage for a specific agent

Use the agent detail page to analyze usage patterns and inspect sessions for an individual agent.

​

Review agent sessions and usage

1. In the Poolside Console, navigate to Agents > Agents.
2. Select an agent.
3. Review the agent usage chart and metrics.
4. Review the list of sessions associated with the agent.
5. Select a session to open it in the Trajectory Viewer.

​

Export usage data for an agent

Export usage metrics for a specific agent to a CSV file.

1. In the Poolside Console, navigate to Agents > Agents.
2. Select an agent.
3. Select the date range for the usage data you want to export.
4. In the Metrics table, click Export Data to download a CSV file.

The export includes summary session counts, token usage, step counts, and inference latency data for the selected agent and date range.

​

Inspect session execution (Trajectory Viewer)

Use the Trajectory Viewer to analyze how an agent executed a session from start to finish. The page displays the user prompt, system instructions, session details, the agent’s final response, and environment information. It also includes Steps, Events, and Time views for inspecting execution flow, detailed events, and time distribution across the session. If the session is still active, the Trajectory Viewer reloads automatically to reflect new activity.

1. In the Poolside Console, navigate to Agents > Agents.
2. Select an agent.
3. Select a session to open it in the Trajectory Viewer.
4. Review high-level session information such as status, start and end time, duration, and whether errors occurred.
5. Review the User Prompt and System Instructions applied during the session.
6. Review the agent’s final response and the Session Exit.
7. Review environment details such as editor, runtime, and platform.

​

Export agent trajectory data

You can export trajectory data as a JSON file to investigate agent behavior or share with Poolside Support for troubleshooting.

1. In the Poolside Console, navigate to Agents > Agents.
2. Select an agent.
3. Select a session to open it in the Trajectory Viewer.
4. Click Download.
5. In the Download Trajectory dialog, review the warning that the file may contain sensitive information.
6. If you want to continue, click Download.

​

Navigate multi-turn sessions

Use the Multi-turn trajectory section to navigate between turns in a multi-turn session.

1. In the Trajectory Viewer, expand Multi-turn trajectory.
2. Review the list of previous turns in the session.
3. Select See turn next to a turn to open the Trajectory Viewer for that turn.

​

Review trajectory steps (Steps tab)

Use the Steps tab to understand the high-level execution flow of a session.

1. In the Trajectory Viewer, select the Steps tab.
2. Review the sequence of steps executed during the session.
3. Identify high-level actions such as reading files, editing content, recording state, or exiting the session.
4. To filter the list to steps with errors, select the Show errors only checkbox.

​

Review trajectory events (Events tab)

Use the Events tab to inspect detailed execution data.

1. In the Trajectory Viewer, select the Events tab.
2. Review detailed execution events, including tool usage, agent messages, and other execution events.
3. Use search to filter events and locate specific details.
4. Inspect individual events to understand how the agent behaved during specific interactions.

​

Review time distribution (Time tab)

Use the Time tab to identify common time sinks within a session by grouping similar tool calls together.

1. In the Trajectory Viewer, select the Time tab.
2. Review tool calls grouped by type to identify patterns across the session.
3. Identify long-running commands, inference-heavy tool calls, and repeated calls for the same inputs.