Observability for local agents

• Define useful traces for agents with tools and RAG.
• Separate technical logs, decisions, and evidence.
• Detect loops, latency, repeated errors, and lack of grounding.

trace:
  request_id: "support-1042"
  user_id: "u_123"
  model: "local-qwen"
  route: "local"
  steps:
    - tool: "search_qdrant"
      latency_ms: 182
      chunks: 5
    - tool: "draft_reply"
      latency_ms: 4210
  guardrails:
    repeated_tool_calls: 0
    human_approval_required: true
  outcome: "draft_created"

What to always log

• Model, runtime, and chosen route.
• Tools called, arguments, and summarized result.
• RAG chunks used and source.
• Repeated errors and strategy changes.
• Whether there was human approval or automatic cutoff.

Official sources

• Langfuse observability overview
• Langfuse tracing
• OpenTelemetry traces