agent-callbacks

Agent observability is achieved by attaching before and after callbacks to LLM model calls and tool execution. These hooks log prompts, available tools, and response summaries, making the agent execution flow transparent and auditable.

Error handling for LLM model calls uses dedicated error callback hooks to intercept failures. These hooks log the issue and can optionally return a recoverable response, allowing production agent orchestration to handle failures gracefully.

Yes, agent execution events can be streamed to Langfuse using external callback handlers. This tracing integration sends model lifecycle, tool execution, and sub-agent delegation events directly to the tracing backend for analytics.

Sub-agent delegation events are captured using specific delegation hooks. These hooks intercept and observe events produced by child agents during an AgentTool execution, providing visibility into nested agent workflows and tool arguments.

Auditing tool arguments and results during async streaming is handled by tool lifecycle callbacks. These hooks capture tool start and end events, recording the exact inputs and outputs of tool execution for verification and debugging.