What is agent observability?
Agent observability makes an execution explainable from request to result, including relationships between models, tools and errors.
It lets teams trace a weak, costly or unsafe outcome to the step that created it.
Why are ordinary logs insufficient?
Separate logs do not clearly show operation order and causality. An agent can call a model repeatedly, choose the wrong tool or enter a retry loop.
A trace groups these events under one task execution.
Trace in an agent
A trace represents a full task and a span represents an important operation. Parent-child relations connect a model invocation to its tool or connector call.
- Agent trace and span
- Model invocation
- Tool and connector invocation
- Result and error
Tool calls and connectors
Track tool identity, minimal approved input, result, error and duration. A connector marks the boundary with an external system.
Captured data must follow permissions and information classification.
Tokens and latency
Measure token use and latency at task, model and step level. Averages can hide loops or slow tools.
Cost and duration per successful task are more useful than per-call totals.
OpenTelemetry for AI
In August 2026 Google aligned richer agent telemetry with OpenTelemetry generative-AI semantic conventions and expanded agent, model, tool and token attributes.
Agent Runtime can also export gen_ai metrics alongside traces and logs.
Privacy in logging
Prompts and responses should not always be logged. They may contain personal data or organizational secrets.
Gemini Enterprise records message content only when observability settings permit it; otherwise it can be redacted or omitted.
- Data minimization
- Redaction
- Limited retention
- Role-based access
- No secret logging
Observability in production
Dashboards should show success, errors, latency, tokens, tool use and escalation. Alerts should detect meaningful changes and budget breaches.
Failed traces should feed evaluation and remediation rather than remain isolated logs.
Enterprise use
Observability supports operational accountability and audit: which agent called which tool for which identity, and with what result.
Telemetry access itself must be restricted and audited.
Frequently asked questions
How is a trace different from a log?
A trace connects the ordered spans of a task; logs are usually individual events.
Should complete prompts be recorded?
No. Content capture requires explicit privacy, redaction and retention controls.
Which cost metric matters most?
Cost per successful task, including steps and retries.
Sources
- Google CloudGemini Enterprise Release Notes — Agent telemetry aligned with OpenTelemetry7 August 2026 · Release Notes
- Google CloudAgent Runtime: ADK telemetry metrics13 August 2026 · Release Notes
- Google CloudInstrument ADK applications with OpenTelemetryOfficial Documentation