Monitor Your Agents in the Atlas Agent Engine UI

For the complete MongoDB documentation index, see www.mongodb.com/docs/llms.txt

Monitor Your Agents in the Atlas Agent Engine UI

MongoDB Atlas Agent Engine is in Public Preview and intended for evaluation and prototyping purposes only. We do not recommend production deployments during preview and expect to communicate breaking changes with you periodically. To learn more about preview feature considerations, see Preview Features.

Overview

The Monitoring page in the Atlas Agent Engine UI shows usage and health metrics for your project and its workspaces. Use the page to track agent invocations, error rates, token usage, memory activity, and resource consumption over time. The page shows data for the current project only.

Each section of the page shows metrics as tiles and trend cards. A tile shows a single value for the selected time period. A trend card shows how a metric changes over the selected time period. Trend cards group data into 5-minute buckets for the 1h period, hourly buckets for the 24h period, and daily buckets for the 7d and 30d periods.

To learn how to view logs, deployment events, and live workspace health, see the [Monitor the Agent](https://www.mongodb.com/docs/agentengine/manage/monitor.md#std-label-agent-engine-monitor guide. To learn how to debug a specific agent run, see the [Inspect Agent Execution Traces](https://www.mongodb.com/docs/agentengine/manage/inspect-traces.md#std-label-agent-engine-inspect-traces guide.

View Metrics

To view metrics for your project, perform the following steps:

  1. Open the Monitoring page.

In the left navigation bar in the Atlas Agent Engine UI, under Monitor, click Monitoring.

  1. Select a time period.

Select 1h, 24h, 7d, or 30d to select a time period to monitor. The page shows the last 24 hours by default.

  1. (Optional) Filter by workspace.

To view metrics for a single workspace, select the workspace from the workspace filter, or click a row in the By workspace or Compute by workspace table. To return to project-wide metrics, select All workspaces.

The page refreshes its data automatically every 60 seconds.

Workspace View

When you view a single workspace, the page shows the same sections as the project-wide view scoped to that workspace, with the following exceptions:

  • The Networking section is not available.

  • The By workspace and Compute by workspace tables are not available.

Customize the Page

You can choose which trend cards appear in each section, and you can show the Compute and Networking sections, which are hidden by default. You can customize the page in the project-wide view or in a single workspace view.

To customize the page, perform the following steps:

  1. Click Customize.

The period selector and workspace filter are hidden while you customize the page.

  1. Select the content to show.

  2. To show or hide a trend card, toggle the card in its section.

  3. To show the Compute section, select Show Compute.

  4. To show the Networking section, select Show networking.

  5. Click Save.

To discard your changes, click Cancel.

By default, the Usage, Memory, and Policy sections show trend cards for the four metrics with the highest volume. You cannot customize the Health trend card.

Your browser stores your selections, so they persist across sessions in the same browser. Your selections do not apply in other browsers or on other devices.

Metrics Reference

This section describes the metrics in each section of the Monitoring page. Each metric shows the total for the selected time period.

Health

The Health section shows the outcome and duration of agent runs. The section also contains Invocations and Errors trend cards. The Errors trend card appears only when errors occurred in the selected period.

The following table describes the metrics in the section:

Metric Description Unit
Invocations Number of agent runs that completed or failed. Runs that are waiting for approval are not counted. Count
Error rate Number of failures divided by invocations. This value is zero when there were no invocations. Percent
Failures Number of runs that ended in an error. Runs that are waiting for approval are not counted. Count
Latency p50 Median time from the start of a run to completion or error, including time spent waiting for approval. ms or s
Latency p95 Time within which 95 percent of runs finished, from the start of a run to completion or error, including time spent waiting for approval. ms or s

The latency metrics appear only when latency data is available for the selected period. To learn more about approvals, see [Human-in-the-Loop Agent Execution.](https://www.mongodb.com/docs/agentengine/deploy/human-in-the-loop.md#std-label-agent-engine-human-in-the-loop

Usage

The Usage section shows model, tool, and runtime consumption. The following table describes the metrics in the section:

Metric Description Unit
LLM calls Number of calls to language models. Count
Input tokens Number of tokens sent to language models. Count
Output tokens Number of tokens returned by language models. Count
Tool sandbox calls Number of calls to tool sandboxes. Count
Tool sandboxes Number of tool sandboxes used. Count
Workload runtime Total time that agent and tool sandboxes ran during the selected period. Duration
Platform uptime Total time that always-on platform services, including the orchestration engine and memory server, ran during the selected period. This value includes idle time. Duration

A Pod runtime trend card, which combines workload runtime and platform uptime, is available through Customize.

Memory

The Memory section shows agent memory activity. The following table describes the metrics in the section:

Metric Description Unit
Docs written Number of documents written to memory in the selected period. This value is not the current size of memory. Count
Docs retrieved Number of documents retrieved from memory. Count
Embedding tokens Number of tokens used to generate embeddings. Count
Searches Number of memory search calls. Count
Context tokens Number of memory tokens added to agent context. Count
Memories consolidated Number of memories added through consolidation. Count

To learn more about agent memory, see [Agent Memory.](https://www.mongodb.com/docs/agentengine/add-features/memory-types.md#std-label-agent-engine-memory

Policy

The Policy section shows the Policy denials metric, which counts the calls that the [Policy Engine](https://www.mongodb.com/docs/agentengine/manage/governance/policy-engine.md#std-label-agent-engine-policy-engine denied.

The metric counts denials from the AUTHORIZED_TOOLS policy type, and from execution and session budget policies. The metric does not count denials from the AUTHORIZED_MODELS policy type. To learn more, see [Policy Types.](https://www.mongodb.com/docs/agentengine/manage/governance/policy-engine.md#std-label-agent-engine-policy-types

Memory Latency

The Memory latency section shows the performance of memory operations. Each metric appears only when your agents performed the related memory operation during the selected time period.

The following table describes the metrics in the section:

Metric Description Unit
Context latency p50 Median time to retrieve memory context. ms or s
Context latency p95 Time within which 95 percent of memory context retrievals finished. ms or s
Write latency p95 Time within which 95 percent of memory writes finished. ms or s
Retrieval success Percentage of memory retrievals that succeeded. Percent

Compute

The Compute section shows resource usage. The section is hidden by default. To show the section, select Show Compute after you select Customize. The section includes CPU and Memory trend cards.

The following table describes the metrics in the section:

Metric Description Unit
CPU Mean CPU usage over the selected period. Cores
CPU user Mean CPU usage in user mode. Cores
CPU system Mean CPU usage in system mode. Cores
Memory Current memory usage. Bytes
Disk Current disk usage. Bytes
Processes Current number of processes. Count

Networking

The Networking section shows project-wide egress traffic. The section is hidden by default and is available in the project-wide view only. To show the section, select Show networking after you select Customize.

Metric Description Unit
Egress throughput Mean outbound traffic rate. Bytes per second
Egress packets Mean outbound packet rate. Packets per second
Active peers Current number of active peers. Count
Peer registrations Number of peers registered. Count
Peer removals Number of peers removed. Count
Stale peers reaped Number of inactive peers removed automatically. Count

When traffic uses more than one IP family, the section also shows a By IP family table with the throughput and share of traffic for each family. To learn more about egress, see [Manage Network Egress Policies.](https://www.mongodb.com/docs/agentengine/network-egress.md#std-label-agent-engine-network-egress

Breakdowns

The Breakdowns section shows how usage is distributed. The section includes the following tables:

  • By model: Token usage and share for each model.

  • By pod type: Total runtime for each pod type and its percentage of the combined runtime of all pod types. The orchestration engine and memory server accrue runtime even when idle.

  • By tool: Number of calls and error rate for each tool.

  • Compute by workspace: CPU usage and share for each workspace. This table is available only in the project-wide view when you show the Compute section.

  • By workspace: Token usage and share for each workspace. Deployed workspaces with no usage show 0. This table is available only in the project-wide view.

Empty and Error States

The page shows the following messages when data is not available depending on the cause:

  • No metrics in this time range: No metrics exist for the 1h or 24h period. Select a longer period, or invoke an agent.

  • No metrics yet: No metrics exist for the 7d or 30d period. Metrics appear after agents run in the project.

  • No data in this time range: A section or table has no data for the selected period. Select a longer time period, or invoke an agent that uses the selected feature.

  • Monitoring data isn't available in this environment yet: Monitoring is not configured for your environment. To learn more, see [Monitor the Agent.](https://www.mongodb.com/docs/agentengine/manage/monitor.md#std-label-agent-engine-monitor

  • Monitoring data couldn't be loaded: The page or a section was unable to load data. The page retries automatically.