Observability & Telemetry

Agno-Go exposes several layers of runtime visibility so that production operators can monitor reasoning-heavy workloads, token consumption, and long-running sessions.

AgentOS SSE Stream

POST /api/v1/agents/{id}/run/stream returns a Server-Sent Events (SSE) stream. Use the optional types query parameter to filter events (e.g. types=run_start,token,reasoning,complete).

Event Types

Event	Description
run_start	Input payload and session metadata.
token	Individual streaming tokens returned by the model.
tool_call	Tool execution metadata (name, arguments, result).
reasoning	Reasoning snapshots with content, token counts, redacted text, model/provider IDs.
complete	Final output, duration, aggregated usage (prompt/completion/reasoning tokens).
error	Structured error object with code and details.

Reasoning events are emitted whenever the model returns ReasoningContent. This works out of the box for OpenAI o1/o3/o4, Gemini 2.5 Thinking, and Claude models with thinking enabled.

Logfire Integration

cmd/examples/logfire_observability demonstrates how to export traces to Logfire using OpenTelemetry. Highlights:

1. Configure the OTLP endpoint and write token (LOGFIRE_WRITE_TOKEN, LOGFIRE_OTLP_ENDPOINT).
2. Build with the logfire tag to include the OpenTelemetry exporter:

   go run -tags logfire cmd/examples/logfire_observability/main.go

3. The example records span metadata such as loop count, aggregated token usage, and reasoning excerpts (with optional redacted content).

Full step-by-step instructions live in docs/release/logfire_observability.md.

Next Steps

• Forward SSE events to your telemetry backend (Logfire, Elastic, Datadog).
• Combine reasoning token metrics with cost dashboards to monitor spend.
• Use the OpenTelemetry hooks to instrument tool execution or end-user request latency.