Managing deployed agents

Consolidated monitoring view of deployed agents, with performance, health, feedback, and consumption metrics, plus execution traces for individual runs.

The Deployed agents tab provides a consolidated view of all deployed agents across all agent types, along with real-time performance, health, and feedback metrics. It acts as the primary workspace for monitoring operational efficiency, diagnosing issues, and ensuring that deployed agents deliver consistent, high-quality outcomes.

Unified agent visibility

This view aggregates all production-ready agents across your tenant, regardless of type, allowing administrators and developers to quickly assess operational status and recent activity. Each agent card displays the agent type (Conversational, Autonomous, or Coded) as a badge, along with key operational metrics:

• Success rate – The percentage of successful runs, helping you identify degraded or underperforming agents.
• Average response time – A measure of execution speed and efficiency across recent jobs.
• Feedback volume – The number of user or system evaluations collected for the agent.
• Health indicators – Visual cues (for example, Healthy, Degraded) reflecting overall performance trends and runtime stability.

Agents can be filtered by state, folder, or time range. Use the Type filter to narrow the view to a specific agent type: Conversational, Autonomous, or Coded.

Operational insights and performance tracking

The tab provides contextual analytics that extend beyond simple runtime status. You can explore:

• Active and completed jobs – Separate timelines track live executions versus completed runs, helping identify workload peaks and bottlenecks.
• Error and latency patterns – Performance charts display execution errors, time-to-first-response, and latency percentiles, allowing teams to fine-tune responsiveness.
• Feedback trends – Aggregated sentiment analysis (positive vs. negative) shows how agent quality evolves over time and which updates affect user satisfaction.

These insights help diagnose systemic issues, validate new deployments, and prioritize optimization efforts.

Note:

Instance management charts are unavailable when CMK encryption is enabled for traces. Because trace attributes are encrypted in Insights and therefore cannot be queried, the analytics required to generate these charts cannot run.

Usage and consumption analysis

The dashboard highlights how agents interact with their environments and how resources are consumed:

• Agent unit usage – Tracks total consumption and identifies outliers in cost or resource utilization.
• Top capabilities invoked – Lists the most frequently used actions (tools, activities, or indexes) across agent runs, helping teams understand behavior patterns and common dependencies.

Detailed agent runtime view

Selecting an individual agent opens a detailed analytics workspace, combining:

• A runtime diagram that maps the agent’s logic, tools, and memory sources.
• Health score visualizations that assess prompt design, tool integration, and evaluation coverage.
• Execution traces with granular visibility into every run (inputs, outputs, errors, and timing details), allowing developers to reconstruct and debug the agent’s reasoning process.

The agent instance view provides full observability into individual agent executions, offering a complete trace of decisions, actions, and performance metrics within a single run.

Understanding the trace view

The trace view combines a graphical representation of the agent’s logic with time-aligned execution data, letting you inspect the flow and outcomes of each step. The canvas shows nodes representing major agent components like models and tools, connected in sequence.

Once you're in the Trace view of an agent run, follow these steps to interact with and investigate the trace data:

1. Preview node information. Point to any node on the canvas to see a tooltip that includes:
   • The execution status (success, retry, or failure)
   • Start and end timestamps
   • A preview of the input and output data This is helpful for getting a quick sense of what happened at each step without needing to open the full details.
2. View full node details. Select any node to open a detailed panel which includes:
   • Complete input and output payloads (JSON)
   • Execution logs or error messages (if applicable)
   • Runtime metrics, such as latency and token usage
   • Configuration parameters, if available
3. Navigate between related steps, to explore how information flowed through the agent:
   • Select connected nodes to move upstream or downstream.
   • This helps you trace how decisions, inputs, or outputs from one step influenced the next.
4. Use the Execution Trace panel. This panel shows a chronological list of all recorded actions during the agent run.
   • Select a row in the trace log to highlight the corresponding node on the canvas.
   • Similarly, selecting a node on the canvas automatically highlights the relevant row in the trace log.