Observability and Instrumentation

Agent objective

Produce a builder with observability configured at the right verbosity level, with optional live streaming and file export, so agent execution can be monitored and debugged.

When to load this skill

• Debugging unexpected agent behavior in development
• Capturing structured logs for post-hoc analysis
• Streaming live execution traces to a dashboard or log aggregator
• Auditing all tool calls and model decisions in production
• Comparing model I/O before/after a system prompt change

Implementation baseline

import { ReactiveAgents } from "@reactive-agents/runtime";

const agent = await ReactiveAgents.create()
  .withName("monitor")
  .withProvider("anthropic")
  .withReasoning({ defaultStrategy: "adaptive", maxIterations: 10 })
  .withTools({ allowedTools: ["web-search", "http-get", "checkpoint"] })
  .withObservability({
    verbosity: "normal",    // show metrics dashboard on completion
    live: true,             // stream events as they happen
    file: "./logs/agent.jsonl",  // write structured JSONL log
  })
  .withAudit()              // record all tool calls and decisions
  .build();

Verbosity levels

Level	Output
"minimal"	No output except final result — for programmatic/embedded use
"normal"	Metrics dashboard on completion (default) — recommended for production
"verbose"	Step-by-step phase summaries as the agent runs
"debug"	Full phase traces including tool call args/results and model responses

.withObservability({ verbosity: "minimal" })   // silent — result only
.withObservability({ verbosity: "normal" })    // dashboard on finish (default)
.withObservability({ verbosity: "verbose" })   // running commentary
.withObservability({ verbosity: "debug" })     // everything, including prompt/response dumps

Key patterns

Live streaming

.withObservability({ verbosity: "verbose", live: true })
// Streams log events in real-time as each phase completes.
// Without live: true, output is buffered and printed at the end.
// Combine with verbosity: "verbose" or "debug" for full traces.

JSONL file export

.withObservability({
  verbosity: "normal",
  file: "./logs/run-2026-04-09.jsonl",   // appends structured events as JSONL
})
// Each line is a JSON object: { timestamp, event, phase, data }
// Suitable for ingestion into log aggregators (Datadog, Loki, etc.)

Model I/O logging

.withObservability({
  verbosity: "debug",
  logModelIO: true,    // log full system prompts and model responses
})
// logModelIO defaults to true at "debug" verbosity, false at all other levels.
// Set logModelIO: false at "debug" to debug phases without exposing prompt content.

Audit trail

.withAudit()
// Records structured audit events for every tool call, guardrail check, contract
// validation, and cost tracking decision.
// Audit events appear in the observability stream and are written to the file if configured.
// Use alongside .withObservability() to capture audit events to a file.

Minimal production config (observability without noise)

.withObservability({ verbosity: "minimal" })
// Disables all terminal output — the agent runs silently.
// Results are returned programmatically only.
// Combine with .withCostTracking() to still enforce budgets without logging.

Development debug config

.withObservability({
  verbosity: "debug",
  live: true,
  logModelIO: true,
  file: "./debug.jsonl",
})
// Maximum visibility: live stream + full model I/O + JSONL file

ObservabilityOptions reference

Field	Type	Default	Notes
verbosity	"minimal"|"normal"|"verbose"|"debug"	"normal"	Output detail level
live	boolean	false	Stream events in real-time
file	string	—	JSONL log file path (appends)
logModelIO	boolean	true at debug, false otherwise	Log full prompts and responses

Pitfalls

• verbosity: "normal" prints a metrics dashboard at completion — this is terminal output, not a structured event. Use file for structured capture
• live: true at verbosity: "debug" produces very high-volume output — only use for targeted debugging sessions
• logModelIO: true logs full system prompts — ensure logs are stored securely, as they may contain sensitive system prompt content
• JSONL file output appends to existing files — rotate or clear the file between runs in long-running test suites
• .withAudit() alone does not produce console output — combine with .withObservability() to see audit events
• At verbosity: "minimal", even errors are not printed to console — check the returned AgentResult for failure details