AI Agent 多步骤任务链路可观测性设计提示词

You are an expert in AI agent observability and distributed tracing.

Design a comprehensive observability strategy for a multi-step AI agent task execution pipeline with the following requirements:

Agent Architecture

• Agent framework: [e.g., LangChain / CrewAI / OpenAI Agents SDK]
• Number of agents: [e.g., 3-5 collaborating agents]
• Task types: [e.g., research, code generation, review]
• LLM providers: [e.g., OpenAI, Anthropic, local models]

Observability Requirements

1. Distributed Tracing: Design trace/span hierarchy for multi-agent task flows
2. Token Tracking: Per-agent, per-step token consumption with cost attribution
3. Latency Profiling: Identify bottlenecks across LLM calls, tool invocations, and inter-agent communication
4. Error Correlation: Link failures across agent boundaries with root cause context
5. Quality Metrics: Track output quality scores, hallucination detection rates, and task success rates

Deliverables

• OpenTelemetry-compatible instrumentation plan
• Grafana/Prometheus dashboard JSON template
• Alert rules for anomaly detection (latency spikes, error rate, cost overrun)
• Structured logging schema with correlation IDs
• Sample code for instrumenting the agent framework

Provide production-ready configurations with clear comments explaining each metric and threshold.