Agent, model, and tool spans

An agent trace starts when Laravel AI dispatches PromptingAgent or StreamingAgent. The root span uses the SDK invocation ID for both the trace ID and root span ID.

Model spans#

For non-streamed responses, Laravel AI SDK 0.10.1 exposes individual provider steps only on the completed agent response. AI Observatory reconstructs one model span for each response step.

These spans contain:

• Provider and model.
• Step number.
• Response text.
• Finish reason.
• Tool calls returned by that model step.
• Token usage.

Because no model-start event was dispatched, reconstructed model spans use the completion observation time for both start and end. Their metadata contains:

{
    "timing": "reconstructed_at_completion"
}

Do not interpret a zero duration on these spans as provider latency.

Tool spans#

InvokingTool starts a tool span and ToolInvoked completes the same span using the SDK tool invocation ID. MCP tools receive the mcp type; other tools receive the tool type.

Tool exceptions may prevent the completion event from being dispatched. Stale recovery handles spans left running by that SDK limitation.