MELT (Metrics, Events, Logs, Traces)

What it is

MELT is the standard shape of observability data in modern software — originated in traditional systems monitoring and now the foundation of AI agent telemetry as well. Metrics are aggregated numbers (latency, error rate, cost, token usage). Events capture significant state changes (a tool call failed, an agent escalated, a handoff happened). Logs record per-step contextual data (the exact prompt sent, the tool response received, the reasoning recorded). Traces link every step of a single request or agent run into one causal chain so you can see cause-and-effect from input to output.

Why it matters

When you instrument AI agents, you are emitting MELT data extended with AI-specific signals — token usage, hallucination scores, drift detection, agent decision paths. Understanding the MELT model makes every observability tool click: OpenTelemetry, Datadog, Salesforce Agent Observability, LangSmith, Arize — they all speak this language. Picking a tool without understanding MELT leads to partial instrumentation and blind spots in production.

Key components

• Metrics — aggregate numeric signals over time (p95 latency, error rate, cost per run, tokens per hour)
• Events — discrete occurrences of significance (tool failure, escalation, deploy, config change)
• Logs — structured contextual data attached to a specific moment (the full prompt, the tool response body)
• Traces — a causal chain of operations that together made up one logical request or agent run

How it works

1. Instrument the system to emit MELT data at every meaningful boundary (function calls, tool invocations, LLM calls)
2. Send the data to an observability backend that organizes it by trace ID, time, and service
3. Use metrics for trends and alerts, traces for debugging single runs, logs for deep context, events for state awareness
4. For AI agents, extend the standard MELT payload with token counts, model parameters, hallucination flags, and quality scores