Intelligent operations
Platform capabilities
Security analytics
AI
Dedicated compute
Dedicated compute keeps concurrent queries stable and complex analysis fast.
Learn moreDedicated compute
Dedicated compute keeps concurrent queries stable and complex analysis fast.
Learn moreContact us
Join the community
Try Guance
Start online with usage-based pricing and a true cloud service.
Get startedChoose a Guance edition
AI Agent observability focuses not only on whether the model interface is successful, but also on the relationships among prompts, completions, models, tools, tools, retrieval, Guardrail, and business chains within a AI Agent session. Guance put Session, Trace, Span, Token, time, and risk events into a single perspective, helping teams see what happens at every step of the AI Agent.
Establishing an observation context around multi-turn conversations, model calls, tool execution, token cost, and risk events, covering Agent/LLM access frameworks like OpenClaw, tool-invoking AI Agent workflows, and enterprise-custom AI Agent applications.
OpenClaw
Hermes
Claude Code
Codex
Other AI AgentAI Agent Execution Process Black Box:A single answer may go through multiple rounds of reasoning, tool calls, retrieval, and Guardrail, making it difficult for traditional links to explain what happens at every step.
Token cost and latency are difficult to control:Models, context length, tool retrys, and multiple rounds of calls all affect cost and response time, and without fine-grained data, optimization is difficult.
Risk events are difficult to locate:Sensitive words, content moderation, tool timeouts, and exception outputs need to be considered together with specific sessions, traces, and spar.
Disconnect between AI applications and business systems:AI Agent calls are only part of the business process and must be analyzed together with logs, metrics, links, and user behavior.
Session and Trace Tracing:Reconstruct the AI Agent execution process by session, turn, and call link, quickly identifying the slow, error, or high performance stages.
Model and Tool Call Analysis:Unified observation of model requests, tool execution, retrieval, time consumption, status, and risk level.
Token and Performance Cost Monitoring:Continuously analyzes token consumption, number of spans, trace time, and model call ratio to support cost optimization.
Risk event linkage positioning:Associate sensitive words, Guardrail, content moderation, and exception states with the specific call context.
The viewer supports filtering data by application, risk level, token range, session ID, and trace ID. Teams can first look at total sessions, high-risk quantity, average number of traces, token consumption, and risk events, then drill into specific sessions or individual traces.
The call analysis view brings tool execution records, risk events, rule hits, status, and risk levels into the same panel. When encountering sensitive content, output anomalies, tool timeouts, or Guardrail hits, teams can directly locate associated Span and call context.
Trace details show the time and order of model requests, tool calls, searches, formatting, and other Span processes in waterfall charts. For multi-turn reasoning, toolchain orchestration, or RAG scenarios, the team can see where the bottleneck is and continue to review input, output, and attribute information.
Through Agent monitoring applications and LLM monitoring applications, teams can generate application IDs, service addresses, and Client Tokens for AI Agent applications. After the application is created, links, logs, metrics, and call data can be aggregated by application dimension, so subsequent troubleshooting no longer relies solely on scattered logs.
LLM observability focuses more on model calls, prompts, tokens, response time, and the application chain of large models; AI Agent observability further revolves around AI Agent sessions, multi-turn traces, tool execution, risk events, and call analysis, suitable for AI applications with tool orchestration, retrieval, Guardrail, or multi-step inference.
Common metrics include session count, number of traces, trace time, number of spans, token consumption, model call ratio, tool execution ratio, number of risk events, risk level, and call status. Teams can use these metrics to assess experience, stability, cost, and security risks.
Teams can create Agent monitoring applications or LLM monitoring applications Guance, obtain application IDs, service addresses, and Client Tokens, and complete configuration according to the access documentation. After data is reported, the viewer analyzes and runs the process by Session, Trace, Token, risk level, and application dimension AI Agent the process.
It is suitable for R&D, algorithm, AI platform, SRE, and security governance teams, especially those already operating AI Agent, customer service assistants, operations Agent, RAG applications, or tool-based AI workflows.
Starting from model requests, prompts, response time, errors, and costs, establish observable analytical entry points for large model applications.
Learn about Agent monitoring applications, LLM monitoring applications, session lists, trace lists, and risk event analysis.
Organize logs, metrics, links, events, RUM, and infrastructure queries into executable, auditable diagnostic processes.
Connect AI IDEs, desktop clients, and automated Agent via Model Context Protocol to make recommendations based on authorized observation data.
Want to see how your AI Agent call chain is fully observed?
Contact sales
