Literature: AgentOps Integration (ADK)

AgentOps provides production-grade observability for autonomous agents, offering session replays, hierarchical tracing, and cost/latency tracking.

🛠️ ADK Integration Strategy

AgentOps employs a "patch and wrap" strategy to provide seamless observability for the Agent Development Kit (ADK).

1. Neutralizing Native Telemetry

AgentOps detects ADK and patches its internal OpenTelemetry tracer (trace.get_tracer('gcp.vertex.agent')) with a NoOpTracer. This prevents duplicate traces and ensures AgentOps remains the authoritative source.

2. Hierarchical Span Mapping

AgentOps wraps key ADK methods to create a logical parent-child relationship:

• Agent Spans (adk.agent.<ClassName>): Parent spans created when run_async starts.
• LLM Spans (adk.llm.<model_name>): Child spans created for model calls. Captures prompts, parameters, and token usage via _finalize_model_response_event.
• Tool Spans (adk.tool.<tool_name>): Child spans created for tool executions. Captures inputs and returned results.

3. Attribute Extraction

AgentOps reuses ADK's internal data extraction logic (patching functions like trace_tool_call and trace_call_llm) to attach rich metadata as attributes to the active AgentOps span.

🚀 Getting Started (Python)

import agentops
import os

agentops.init(
    api_key=os.getenv("AGENTOPS_API_KEY"),
    trace_name="my-adk-trace" # Optional
)
# AgentOps now automatically instruments all ADK Runner and Agent calls.

📊 Visualization Features

• Waterfall of Spans: Displays the sequence and duration of nested sub-agent and tool calls.
• Session Replay: Allows developers to re-watch the agent's decision-making process.
• Cost Tracking: Aggregates token usage across multiple providers into a single dollar-denominated metric.

References

• Source: 00_Raw/adk-documentation.md (Lines 16087-16209)
• agent-observability
• adk-moc
• lit-otel-genai