Viewing Agent Observable Application Details
Download
Focus Mode
Font Size
Note:
If you are using an AI tool, you can install the Tencent Cloud Agent Observability Access Assistant Skill to enable the AI to complete observability access and data analysis for Agent applications in one stop. After access is completed, you only need to describe your analysis requirements in the AI conversation (for example, "help me check the recent operational status of the Agent"), and a multi-dimensional analysis will be automatically performed based on the reported Trace data.
Overview
The Application Details page is the core feature page for Agent observability, providing a comprehensive monitoring and analysis view for integrated AI applications. Go to CLS Console > Agent Observability to view application details.
The page contains three major feature modules: Dashboard, Call Chain, and Session, helping you comprehensively grasp the operational status of AI applications from macro metrics to micro-level links.
Scenario | Capability | Benefit |
What is the overall operational status? | Dashboard - Overview | Provides an overview of core metrics such as number of requests, number of errors, number of model invocations, and Token consumption, and supports Top10 rankings by Agent/model dimension. |
Whether performance has a bottleneck. | Dashboard - Performance | Request latency distribution, P50/P90/P99 percentile trends, model average latency ranking, and identification of slow models and long-tail delays. |
Where Token Costs Are Incurred | Dashboard - Cost & Token | Input/Output Token totals and percentile trends, enabling users to grasp cost trends. |
How Internal Application Components Perform | Dashboard - Application Observability | Analyze invocation distribution and latency by operation type (generation/tool/agent/retriever). |
What Happens in a Single Request | Call chain | Provides a complete Trace list and details, supports filtering by status/error type/latency, and displays the call tree and the Input/Output of each node. |
Complete Picture of User Multi-turn Dialog | Session | Aggregates Traces by Session dimension, restores multi-turn conversation context, and calculates session-level Token and cost. |
Application List
Go to the CLS Console > Agent Observability page to view all integrated AI applications. The top of the page displays the following global metrics:
Metric Value | Description |
Applications Connected | Total number of applications that have been successfully connected to the observable Agent. |
Applications Reporting Data | Number of applications that are actively reporting data. |
Total Data Written Yesterday | Total volume of data written to CLS by all applications on the previous day. |
Total Tokens Yesterday | Total number of Tokens consumed by all applications on the previous day. |
The application list displays information for each application, such as its name/ID, region, instance status, access type, yesterday's write data volume, and yesterday's Token quantity. Click Enter Observation in the Actions column to go to the details page for that application.
Dashboard
The Dashboard page provides a visual monitoring panel for application operations, intuitively displaying various core application metrics through metric cards and time-series charts. The dashboard contains four sub-tabs: Overview, Performance, Cost & Token, and Application Observability.
Overview
The Overview panel provides a global perspective on application operational status, making it suitable for daily inspections and rapid anomaly identification.
Function | Description |
Core Metrics Card | Displays core KPIs such as number of Agents, number of requests, number of model invocations, input/output Tokens, and total Tokens, and shows the month-over-month change percentage. |
Request/Error Quantity Time Series | Shows the trend of request volume and number of errors over time, enabling rapid detection of sudden spikes and drops. |
Model Invocation Quantity Time Series | Time series variation of underlying large model invocation volume. |
Top 10 Request Quantity (by Agent) | Ranks request volume by Agent and identifies frequently invoked Agents. |
Top 10 Model Invocation Quantity (by Agent + Model) | Ranks invocation volume by the combined dimension of Agent and model. |
Total Tokens Distribution | Histogram distribution of Token consumption per request, showing the concentration interval of Token usage. |
Top 10 Total Tokens (by Agent/Model) | Identifies high-cost Agents and high-consumption models. |
Model Invocation Error Count Time Series | Observes whether model errors exhibit regular fluctuations. |
Tokens Per Second Output (by Model) | Evaluates the output throughput rate of each model. |
Average Time to First Token (TTFT) | Reflects the user-perceived response speed. |
Average Number of Model Invocations/Tokens Consumed per Request (by Agent) | Understands the model invocation complexity and per-call cost of each Agent. |
Performance
The Performance panel focuses on latency and throughput analysis, helping to identify and optimize performance bottlenecks.
Function | Description |
Core Metrics Time Series | Trends of request volume, number of errors, average latency, number of model invocations, number of model errors, and model average latency. |
Request Latency Distribution | A histogram of request latency, with the P50 percentile line marked, to intuitively show the concentration interval of latency. |
Request Latency Percentiles | Time-series trends of P50, P90, and P99 latency for monitoring long-tail delay variations. |
Model Invocation Latency Distribution | Distribution of latency histogram at the model level. |
Model Average Latency Time Series | Shows the variation of average latency per model over time to detect model performance degradation. |
Top 10 Model Invocation Quantity / Average Latency | Ranks invocation volume and latency by model to identify slow-performing models. |
Top 10 by Agent + Model Combination | Ranks request volume and number of model invocations by the combined dimension. |
Cost & Token
The Cost & Token panel helps you monitor and manage Token consumption trends.
Function | Description |
Input/Output Tokens | Total number of input and output tokens within the selected time range. |
Total Tokens Time Series | Variation trend of input and output Token counts over time. |
Token Percentile Time Series | Time-series trends of P50/P90/P99 percentiles for Token consumption, identifying requests with abnormally high consumption. |
Application Observability
The Application Observability panel provides operational analysis at the application's internal component level, displaying the behavior and performance of each component by the dimension of operation type (Observation Type).
Function | Description |
Operation type distribution | Displays the invocation percentage of each operation type, such as generation, guardrail, tool, chain, agent, retriever, and event. |
Observation average latency | Compares the average execution latency of each operation type to identify slow components. |
LLM Invocation/Error Count Time Series | Variation trend of LLM invocation quantity and error quantity over time. |
LLM Latency Percentile Time Series | Time-series trends of P50/P90/P99 latency for LLM invocations. |
LLM Model Invocation Ranking | Ranking of LLM models by number of invocations. |
LLM Average Latency Top 10 | Ranks LLM models by average latency to identify slow models. |
Call-Chain
The Call Chain page provides request-level, fine-grained tracing capabilities, helping you deeply analyze the complete execution path of a single request. It supports switching between the Traces and Observations views.
Trace List and Filtering
Function | Description |
Filter by Trace Status | Filter by success (OK) or error (ERROR) status to quickly focus on abnormal requests. |
Filter by Error Type | Filter by specific error types (tool invocation failure, LLM invocation failure, Root span status code exception, AGENT execution failure, and so on). |
Query by Trace ID / Session ID | Precisely locate a specific Trace or all Traces under a session. |
Filter by Duration Range | Filter by duration range to quickly locate slow requests. |
Displaying the List | Displays status, error type, timestamp, Trace ID, Operation name, invoked model, and other information for each Trace. |
Trace Details
Click any Trace to expand the details view, which fully restores the call chain for a single request.
Function | Description |
Basic Information | Trace ID, start time, total duration, number of Observations, total Token consumption (input→output), cost, version number, Session ID. |
Call tree | The left panel displays the complete call chain of this request in a tree structure, including the operation type (generation/tool/guardrail), duration, and error status indicator of each node. |
Node details | Select any node in the call tree, and the right panel displays the node's Tags, Input/Output content, Observation type, latency, and execution status. |
Error Diagnosis | Traces in ERROR status are annotated with the source of the error signal (for example, "triggered by Observation errors that bubbled up to the Trace"), helping to quickly locate the root cause. |
Session
The Session page aggregates and displays the interaction process between users and AI applications by session dimension, helping you understand user usage patterns and conversation quality.
Function | Description |
Search by Session ID | Supports precise search for specific sessions. |
Filter by Session Duration | Filters sessions based on the session duration range. |
Filter by Traces Count | Filters sessions based on the number of Traces they contain. |
Filter by Input/Output/Total Tokens | Filters sessions based on Token consumption at the session level. |
Filter by Input/Output Cost | Filters sessions based on cost at the session level. |
Session List | Displays information such as Session ID, creation time, duration, model, and user input summary. |
Session Details | After a session is selected, displays all Trace call trees under that Session, and the right panel displays Trace details (Tags, Input/Output content, Token consumption, cost). |
Help and Support
Was this page helpful?
You can also Contact sales or Submit a Ticket for help.
Help us improve! Rate your documentation experience in 5 mins.
Feedback