Grafana Labs為可觀測性平台新增 AI 監測智能體與程式碼修復智能體

Grafana Labs is expanding Grafana Cloud with AI agents that watch for emerging problems, tools that prepare code fixes for human review, and a shared knowledge graph connecting infrastructure failures to the users they affect. Announced at ObservabilityCON 2026, the releases address two increasingly connected jobs: using AI to operate software and monitoring the AI systems entering production.

The rollout combines generally available capabilities with public and private previews. Assistant Watchers and Agent Observability with OpenTelemetry support are generally available, while Grafana Cloud Agents enters public preview. Alongside them, Grafana is introducing faster log searches, additional telemetry cost controls, and a commercial plugin marketplace.

The underlying question is how far an operations team should let an agent go. Finding an unusual pattern, explaining a failure, and changing production code carry different consequences. Grafana’s announcement brings those steps closer together while retaining human approval for proposed code fixes.

From watching telemetry to proposing a fix

Assistant Watchers lets teams describe what they want monitored in ordinary language. Rather than relying exclusively on a predefined threshold, a Watcher learns patterns from telemetry and surfaces evidence when those patterns change. Findings route through existing Grafana Alerting, and a critical finding can initiate a deeper investigation.

That extends Grafana Assistant, the company’s built-in agent for asking questions about operational data. Assistant Investigations, generally available since July, can perform longer investigations initiated by alerts or incidents. The new Watchers supply another starting point: an emerging change that a team may not yet have encoded as an alert rule.

Grafana Cloud Agents adds a controlled cloud environment where Assistant can examine repository code alongside live telemetry, prepare a proposed fix, and open a pull request. The public preview stops at a reviewable change for a person to approve. A plausible diagnosis and a working patch still need validation against the application’s requirements and deployment process.

This matters because an agent can produce a convincing explanation without having enough evidence. Connecting each finding to telemetry gives engineers something to inspect, while a pull request creates a familiar checkpoint for reviewing the actual change.

Watching the agents themselves

The other half of the announcement concerns AI workloads already running inside applications. Agent Observability’s generally available OpenTelemetry support brings agent activity into the same open instrumentation standard used for other services.

Grafana’s Agent Observability product documentation describes conversation replay across prompts, tool calls, and responses, along with latency, token consumption, model costs, evaluations, and experiments. That can help teams distinguish an expensive model call from a slow downstream service, or investigate a quality regression after a prompt or tool changes.

Infrastructure health and answer quality require different measurements. A service can remain available while an agent uses the wrong tool or returns a poor response. Combining operational telemetry with evaluations broadens what engineers can investigate; it does not make every quality judgment automatic.

A knowledge graph that connects symptoms to impact

Grafana is also placing a Knowledge Graph in every Grafana Cloud stack. The runtime context layer links applications, databases, cloud services, Kubernetes infrastructure, and user-facing telemetry. Teams can add custom entities and relationships to represent the products or business systems that matter to them.

Custom entities and runtime context for agents are generally available. A new Cypher query capability allows agents to retrieve relationships from the graph, and Grafana says that context is accessible to Assistant and agents connected through gcx. The point is to reuse an established view of dependencies instead of reconstructing it from raw logs during every investigation.

Grafana’s existing Cloud MCP server documentation provides useful context for its approach to external agents. The hosted server lets compatible clients query Grafana data using OAuth authorization, with access limited by the user’s Grafana permissions. Sharing operational context across tools therefore also depends on the permissions and boundaries configured around that data.

Several extensions remain previews. A service catalog with ownership, Mobile Observability, Critical User Journeys, and Operational Context with Knowledge Graph are in private preview. Session Replay and Grafana Workflows are in public preview. Response Console, which brings human responders and agents into a common incident workspace, is in private preview.

These distinctions matter for adoption. A team can build on generally available graph capabilities now, but should assess preview access and maturity separately before depending on the broader incident-response workflow.

More telemetry makes cost control part of the AI stack

Agents generate additional activity to inspect: model requests, tool calls, code changes, and the behavior of the applications they touch. Keeping that information useful without retaining every signal at full cost is another focus of the release.

Grafana Cloud Logs gains generally available full-text search enhancements based on technology from Logline, which Grafana acquired in April 2026. Grafana says the changes accelerate searches for rare values such as request IDs over object storage, at no additional cost. This is especially relevant to investigations that need one specific event rather than an aggregate trend.

Adaptive Logs Archive, in public preview, routes logs to a customer-owned object-storage bucket in an open format. Adaptive Metrics Auto-Apply, generally available, automatically applies cardinality-reduction recommendations, with a Cost Mode intended to prevent increases in total active series. Another generally available update lets developers retrieve traces by ID from the previous 24 hours even when Adaptive Traces dropped them.

The tradeoff is operational as well as financial: retention and sampling choices determine what evidence remains available after a failure. Customer-owned archives and a defined trace-retrieval window give teams more options, but they still need policies that fit their investigation and compliance needs.

A commercial marketplace and a measured path to autonomy

The generally available Grafana Marketplace adds paid plugins from verified partners to the existing plugin ecosystem. Paid-tier Grafana Cloud customers can start a 14-day trial and subscribe monthly through their existing invoice. Partners provide plugin support, while Grafana handles billing. Existing free plugin distribution remains available.

The announcements arrive alongside Grafana Labs and ETR’s 2026 State of Observability Report, based on 300 IT and engineering leaders. According to the releases, 68% have moved AI for observability beyond investigation into pilots or production. Comfort is substantially higher for read-only investigation, at 85%, than for fully autonomous remediation, at 52%.

Those survey responses describe adoption and attitudes, rather than demonstrating the reliability of the newly announced tools. They nevertheless explain the shape of this rollout: automate detection and investigation, supply a shared record of dependencies, and make proposed action reviewable. For engineering teams, the practical test will be whether that combination produces better evidence and fewer repeated investigations while keeping responsibility for consequential changes clear.