Grafana adds AI agents to monitor and fix incidents

Grafana Labs enhances Grafana Cloud with AI agents to detect anomalies, prepare fixes, and link incidents, code, and user impact. The publisher also focuses on a shared knowledge graph and new cost and search tools.

Grafana Labs announced several additions to Grafana Cloud at ObservabilityCON 2026: AI agents that monitor weak signals, propose fixes for human review, and link outages to affected users. The whole set addresses a sensitive point for operations teams: automating without letting a machine decide alone on a production change. At the same time, the publisher aims to better observe the AI systems already entering production, a topic far less theoretical than two years ago.

Agents that spot anomalies before they become tickets

Assistant Watchers, now available, allows describing in plain language what needs to be monitored. Rather than sticking to a fixed threshold, the tool learns patterns in telemetry and raises items when these patterns change. Results go through Grafana Alerting, with the possibility to open a deeper investigation if the signal seems critical.

Grafana thus extends its assistant already integrated into the platform. Assistant Investigations, available since July, could already launch long analyses from an alert or incident. Watchers adds a more flexible entry point: an emerging evolution that the team has not yet translated into an alert rule. It’s less glamorous than a sci-fi movie but much more useful at 3 a.m.

Another piece of the setup: Grafana Cloud Agents, in public preview. The controlled cloud environment allows Assistant to examine repository code alongside live telemetry, then prepare a fix as a pull request. Validation remains human. Grafana therefore does not promise a robot that repairs on its own, but a system that prepares the ground and leaves a familiar control point to engineers.

Observing the agents themselves changes the game

The other half of the announcement concerns applications already using models. Agent Observability, with OpenTelemetry support, is available and integrates agent activity into the open instrumentation standard already used for other services. Grafana also describes a conversation replay, which allows replaying prompts, tool calls, and responses, with latency, token consumption, model cost, evaluations, and experiments.

This layer does not make automatic judgments. However, it helps distinguish a model that is too costly from a slow downstream service, or a quality regression after a prompt change. An application can remain available while an agent chooses the wrong tool. In this case, infrastructure metrics are not enough.

To link these signals to what really matters, Grafana also adds a Knowledge Graph in each Grafana Cloud stack. The context layer connects applications, databases, cloud services, Kubernetes infrastructure, and user-oriented telemetry. Teams can add their own entities and relationships, then query this graph via Cypher. The goal is simple: avoid rebuilding the dependency map by hand at each incident.

Cost and data access become part of the debate

Grafana also pushes on volume control. Grafana Cloud Logs receives full-text search improvements, derived from Logline technology, a company acquired in April 2026. The publisher says it accelerates the search for rare values, like a request ID, in object storage, without additional cost. Adaptive Logs Archive, in public preview, sends logs to a client-owned object bucket, in an open format.

On its side, Adaptive Metrics Auto-Apply, available, automatically applies cardinality reduction recommendations, with a cost mode intended to avoid an increase in the number of active series. Grafana also adds trace retrieval over the last 24 hours, even when Adaptive Traces have discarded them. These trade-offs matter because retention, sampling, and archiving determine what remains to be examined after an outage.

The platform finally introduces a commercial Grafana Marketplace, with paid plugins from verified partners. Paying Grafana Cloud customers can test these extensions for 14 days, then subscribe monthly via their existing invoice. Free plugins remain distributed as before, but the publisher opens a clearer path to monetize the module ecosystem.

Ce que disent les chiffres de l’observabilité agentique

The announcements come with a market reading. According to the 2026 State of Observability Report by Grafana Labs and ETR, based on 300 IT and engineering leaders, 68% of respondents have already moved AI for observability beyond simple investigation, towards pilots or production. Usage remains more reassuring when limited to incident reading.

Still according to this report, 85% feel comfortable with AI in read-only mode for investigation, versus 52% for fully autonomous remediation. This difference sums up the sector’s moment: teams want to save time on detection and analysis but keep control as soon as it comes to touching the system. Trust is growing, full delegation much less so.

68% — share of respondents who have moved beyond AI experimentation, according to Grafana Labs and ETR, 2026.

85% — comfort level with read-only AI, according to Grafana Labs and ETR, 2026.

52% — comfort level with fully autonomous remediation, according to Grafana Labs and ETR, 2026.

300 — IT and engineering leaders surveyed for the 2026 report, according to Grafana Labs and ETR.

To read