Agent Logs
Agent Logs provide complete visibility into how an Agent executes a request.
Every execution generates a sequence of structured events that allow teams to understand:
- What happened
- When it happened
- Which tools were used
- Whether approvals were requested
- Whether execution was transferred to another agent
- Why an execution failed
This makes AI systems observable, debuggable, and auditable.
Agent Event Lifecycle
Section titled “Agent Event Lifecycle”A typical execution may look like this:
agent_start ↓tool_call ↓sub_prompt ↓mcp_tool_call ↓hitl_request ↓agent_response ↓agent_completeSupported Events
Section titled “Supported Events”agent_start
Section titled “agent_start”Indicates the beginning of an agent execution.
Example:
agent_startAgent: IT Support AssistantRun ID: agent_01HABC123tool_call
Section titled “tool_call”Generated whenever the agent executes a built-in Skill.
Examples include:
- PostgreSQL
- Web Search
- Email Send
- Vector Search
- SSH
- Local Command
Example:
tool_callTool: PostgreSQLOperation: Employee LookupStatus: Successmcp_tool_call
Section titled “mcp_tool_call”Generated whenever an MCP Tool is executed.
Examples include:
- GitHub Issue Creation
- Jira Ticket Creation
- Slack Notification
- Internal API Access
Example:
mcp_tool_callServer: GitHub MCPTool: Create IssueStatus: Successsub_prompt
Section titled “sub_prompt”Represents internal prompts generated by the agent during execution.
Sub-prompts are commonly used for:
- Planning
- Reasoning
- Tool selection
- Decision making
Example:
sub_promptDetermine whether approval is required for VPN restoration.hitl_request
Section titled “hitl_request”Generated when the agent requires human approval before continuing execution.
Example:
hitl_requestApproval Type: Manager ApprovalStatus: Pendingagent_transfer
Section titled “agent_transfer”Generated when execution is transferred to another agent.
Example:
agent_transferFrom: IT Support AssistantTo: Security AgentReason: Security policy validation requiredagent_response
Section titled “agent_response”Generated when the agent produces a response.
Example:
agent_responseVPN restoration requires manager approval.An approval request has been sent.agent_complete
Section titled “agent_complete”Indicates successful completion of execution.
Example:
agent_completeStatus: SuccessDuration: 8.2sagent_error
Section titled “agent_error”Generated when execution fails.
Examples include:
- Tool failures
- Authentication failures
- Timeouts
- Model errors
- Invalid inputs
Example:
agent_errorReason: GitHub MCP authentication failedExample Execution Timeline
Section titled “Example Execution Timeline”10:01:02 agent_start10:01:03 tool_call10:01:04 sub_prompt10:01:05 mcp_tool_call10:01:06 hitl_request10:01:10 agent_response10:01:11 agent_completeMulti-Agent Example
Section titled “Multi-Agent Example”10:01:02 agent_start10:01:04 agent_transfer10:01:05 agent_start10:01:06 mcp_tool_call10:01:08 agent_response10:01:09 agent_completeWhy Agent Logs Matter
Section titled “Why Agent Logs Matter”Agent Logs provide:
- Execution visibility
- Debugging capabilities
- Auditability
- Governance
- Production observability
They transform AI systems from black boxes into transparent and manageable systems.
Next Steps
Section titled “Next Steps”Continue with:
- Workflow Logs
- RAG Logs
- Execution Traces
- Token Usage
- Cost Analytics