Instrumentation Planning for Agents

Plan agent observability using a tiered, outcome-focused approach.

Core Principle

Every metric and span should answer one of these questions:

1. Did the agent complete its task? (success/failure)
2. How long did it take? (latency)
3. How much did it cost? (tokens/money)
4. Why did it fail? (error context)
5. What decisions did it make? (reasoning trace)

5-Tier Implementation Framework

Tier 0: Foundation (Day 1)

Essential observability to ship any agent:

• SDK initialization
• Root span for agent runs
• Unhandled error capture
• Basic success/failure status

Tier 1: Core Tracing (Week 1)

Understand agent execution:

• LLM call spans (model, latency)
• Tool execution spans (name, result)
• Agent loop iterations
• Retry attempts

Tier 2: Context & Attribution (Week 2)

Track costs and ownership:

• Token counts (input/output/total)
• Cost per call (USD)
• User/session context
• Feature/workflow attribution

Tier 3: Multi-Agent Coordination (Week 3)

For multi-agent systems:

• Parent-child span relationships
• Agent handoff tracking
• Delegation reasoning
• Supervisor decisions

Tier 4: Evaluation & Quality (Month 1)

Measure agent quality:

• Response quality scores
• Human feedback capture
• Automated eval results
• Hallucination detection signals

What NOT to Instrument

• Full prompt/response content (PII, storage cost)
• Every intermediate thought (noise)
• Timestamps as attributes (use span timing)
• User-provided secrets

Span Naming Convention

Use semantic, hierarchical names:

agent.run              # Root agent execution
agent.think            # Reasoning step
llm.call               # LLM API call
llm.stream             # Streaming LLM call
tool.execute           # Tool execution
tool.validate          # Tool input validation
retrieval.search       # RAG retrieval
retrieval.rerank       # Reranking step
memory.read            # Memory fetch
memory.write           # Memory store
handoff.delegate       # Agent delegation
handoff.receive        # Receiving delegation
human.request          # Human approval request
human.response         # Human response received
eval.score             # Evaluation scoring

Attribute Naming Convention

Use dot-notation, consistent types:

# Agent context
agent.name             # string: "researcher"
agent.type             # string: "langgraph"
agent.run_id           # string: UUID

# LLM context
llm.model              # string: "claude-3-opus"
llm.provider           # string: "anthropic"
llm.temperature        # float: 0.7
llm.tokens.input       # int: 1500
llm.tokens.output      # int: 350
llm.tokens.total       # int: 1850
llm.cost_usd           # float: 0.025
llm.latency_ms         # int: 2340

# Tool context
tool.name              # string: "web_search"
tool.success           # bool: true
tool.error             # string: error message
tool.latency_ms        # int: 450

# User context
user.id                # string: hash or ID
session.id             # string: session UUID

Sampling Strategy

For high-volume agents:

• 100%: Errors, slow requests (>P95), human escalations
• 10-25%: Normal successful runs
• 1%: High-frequency background tasks

Configure sampling at SDK level, not in code.

Related Skills

• llm-call-tracing - LLM instrumentation details
• tool-call-tracking - Tool execution patterns
• token-cost-tracking - Cost calculation methods