AgentOps DoctorLive dashboard →

Send your first agent trace

AgentOps Doctor accepts standard OTLP/HTTP traces. If your agent already uses OpenTelemetry, no model-provider code needs to change.

1. Configure the exporter

Set the standard OTLP trace endpoint and include your project ingest key.

shell
export OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=https://your-host/v1/traces
export OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer adk_your_key"

2. Emit GenAI semantic conventions

The analyzer understands OpenTelemetry’s standard GenAI operations including invoke_agent, chat and execute_tool.

shell
span.setAttribute("gen_ai.operation.name", "execute_tool")
span.setAttribute("gen_ai.tool.name", "inventory_check")
span.setAttribute("agentops.tool.timeout_ms", 5000)
Keep sensitive content out by default

Prompt, completion, retrieval and tool payload bodies are not required for core reliability diagnosis. Hash repeatable tool inputs when you want precise loop detection without storing arguments.

Telemetry model

Every run becomes a trace. Agent invocations, model requests and tool calls stay as spans beneath that trace. AgentOps Doctor derives findings from timing, status, operation type, tool identity and usage metadata.

Initial detection rules

  • Tool loop — repeated execution of the same tool/input path in one agent run.
  • Missing timeout — long tool execution with no declared timeout metadata.
  • Provider recovery gap — failed model provider with no later cross-provider fallback.

Privacy

AgentOps Doctor’s core schema stores attributes, status, durations and usage. Teams can omit prompt and response content entirely and still receive reliability findings.

Ingestion headers

shell
Authorization: Bearer adk_...
Content-Type: application/x-protobuf