Grafana Agent Observability is GA: trace your AI agents

If you run an AI agent in production and can't say what it did yesterday or what it cost, you are flying blind. On October 20, Grafana Labs made Agent Observability with OpenTelemetry support generally available. Agent activity, cost and behavior now land in the same dashboards as the rest of your stack.

What actually happened

At ObservabilityCON 2026, Grafana Labs announced four AI updates:

• Agent Observability with OpenTelemetry: generally available. It shows what agents do, what they cost, and whether their behavior holds up over time, without a separate tool.
• Assistant Watchers: generally available. Always-on agents learn normal telemetry patterns, flag changes through Grafana Alerting, and can open an investigation on critical findings.
• Grafana Managed Sandboxes: public preview. Grafana Assistant can clone a repo, prepare a fix, and open a pull request for a human to review.
• Grafana Pages: public preview, for visualization beyond fixed dashboard layouts.

The key design choice is OpenTelemetry. Agents report on the same open standard as your APIs and databases, so an agent shows up as one more service. Grafana's press release cites its own survey: 76% of organizations already invest in OpenTelemetry.

The release does not list pricing or say which agent frameworks have out-of-the-box support. Check Grafana's docs for your stack before you plan around it.

Why it matters for your business

An agent is a service. Monitor it like one. It calls APIs, burns tokens, fails and retries. If your web app has alerts but your agent only has a chat log, you will find out about the bad week from a customer.

Cost per task beats cost per month. Your monthly AI bill tells you nothing. Cost per refund handled, per quote drafted, per ticket closed tells you whether the agent pays for itself. You need traces to get that number.

OpenTelemetry keeps you portable. Instrument with an open standard and you can move from Grafana to Datadog, Honeycomb or a self-hosted stack without re-instrumenting. Proprietary agent-monitoring SDKs lock your history to one vendor.

Read the preview label. Managed Sandboxes lets an AI open pull requests on your code. It is preview, and it still needs a human to approve every merge. Keep it that way.

Sources: Grafana Labs.