Grafana ships autonomous Watchers and AI-prepared fixes in sandboxes
On October 20, 2026, at ObservabilityCON, Grafana Labs made its Assistant Watchers and OpenTelemetry-compatible Agent Observability generally available, and previewed Managed Sandboxes and Grafana Pages. For anyone running LLM agents in production, the stakes run both ways: let AI operate the system, and observe the agents themselves.
October 20, 2026. At ObservabilityCON, Grafana Labs makes its Assistant Watchers and OpenTelemetry-compatible Agent Observability generally available, and previews Managed Sandboxes and Grafana Pages. 2026. The State of Observability report released the same week finds that 68% of the 300 surveyed leaders have moved AI for observability past investigation. September 2026. An IDC white paper pegs mean-time-to-resolution improvements at 63% for customers using Grafana’s AI tools. The point that matters: the shift is no longer about automating dashboards, but about letting agents detect, diagnose, and prepare fixes — with a human holding the reins at every step.
Two problems, one platform
Grafana Labs frames the announcement as the answer to two overlapping needs. The first is using agents to run systems: spot an anomaly, run the investigation, propose an action. The second is understanding what the agents themselves do once they are in production. The company names the two sides “AI for observability” and “observability for AI,” and builds one platform that treats them together.
The report’s numbers expose the tension. 85% of leaders say they are comfortable with read-only investigation, but only 52% are comfortable with fully autonomous remediation. Adoption is real but lopsided: teams happily hand AI the log-reading, but far less the permission to restart a service. And while 80% of organizations are building LLM-powered features, 45% have pieced their AI observability together from open source, versus only 25% running a purpose-built product.
In practical terms, the shape of on-call work changes. Instead of staring at a dashboard hoping to spot the drift, a team describes in plain language what it watches, and the agent flags when a pattern changes — with the evidence attached. The human work shifts from detection to judgment: decide whether the finding deserves an investigation, then approve or reject the proposed action.
The open question is trust. Agents that act on telemetry can be wrong in expensive ways — a false positive that opens a spurious pull request, or a cost run that silently drains the inference budget. Grafana’s answer is structural rather than rhetorical: every autonomous step ends in a human gate, and every agent action is itself telemetry. The bet is that teams grant autonomy in proportion to the evidence they can see, not in proportion to a vendor’s confidence.
Watchers: catching what alerts miss
The centerpiece is the general availability of Assistant Watchers. A Watcher is an always-on agent tasked with spotting changes that no predefined alert rule anticipated. A team describes in plain language what it cares about; the Watcher learns normal patterns from the telemetry and raises a finding when those patterns drift.
Each finding arrives with the evidence behind it, and routes through Grafana Alerting, the existing alert channel. A critical finding can trigger an Investigation — the long-running, automated inquiry Grafana made generally available in late July. The pitch is precise: Watchers do not replace threshold-based alerts, they complement them by catching the slow drifts that never cross a clean threshold.
Managed Sandboxes: the agent that prepares the fix
The most ambitious preview is called Managed Sandboxes. It gives Grafana Assistant a controlled cloud environment to develop fixes, drawing on both repository code and production telemetry. The scenario is concrete: the assistant clones a repository, examines the code in light of observed production behavior, prepares a fix, and opens a pull request for human review.
That is the piece that closes the loop: Watchers detect, Investigations diagnose, Sandboxes prepare the action, and a human approves. Mat Ryer, senior director of AI at Grafana Labs, sums up the line: teams want agents to do more of the work, “but they need evidence and control at each step.”
Agent Observability: seeing agents as services
On the other side of the mirror, Agent Observability reaches general availability with OpenTelemetry support. The reasoning: agents are now production workloads, and they must be visible like any other service — what they do, what they cost, and whether their behavior holds up over time.
The standard is not cosmetic. 76% of organizations already invest in OpenTelemetry, according to the vendor’s 2026 survey. Tying agent activity to that standard means it shows up in the same pipeline as the rest of the stack, with no extra tool, using the same traces, metrics, and logs. For an SRE team, that is the difference between watching your agents and letting them run blind.
Grafana Pages: beyond the dashboard
The final preview, Grafana Pages, addresses an old limitation of the product: a dashboard imposes a grid of panels, whereas an incident sometimes demands a bespoke view, closer to an app than a monitoring screen. Pages lets teams compose visualizations outside any fixed constraints, while reusing existing components. The vendor presents it as “the next evolution of Grafana”: less a dashboard than an open surface, able to host the output of a full investigation.
Governance, meanwhile, is non-negotiable. Investigations can be launched from the platform, Slack, the Grafana Cloud MCP server, or the gcx command-line tool. Whatever the entry point, actions that change a system stay under human control: a Watcher can open an inquiry, not restart a service without approval. That is the line Grafana draws between “AI for observability” and “observability for AI.”
The numbers behind the promise
The marketing rests on dated measurements. The IDC white paper from September 2026 — sponsored by Grafana Labs, based on a four-organization sample — reports a 63% reduction in mean time to resolution and 46% less time spent gathering context during on-call shifts. Directional figures, acknowledged as such by IDC, but they give an order of magnitude.
The field backs the promise. Juan Carlos Ramirez, senior staff DevOps engineer at SailPoint, describes the day-to-day use of Grafana Assistant: ask a question, let the investigation run, and return to a full write-up — the time to find a spike in logs dropping, by his account, from fifteen minutes to five, “while I’m doing other things.”
One caveat worth keeping: the IDC sample is small and vendor-sponsored, so the exact percentages carry less weight than the direction — context-gathering and resolution time both shrink when an agent does the legwork.
Verdict
If you run LLM agents in production, wire their observability into OpenTelemetry now: it is the only way to see what they do and what they cost without adding another tool. If you are ready to let AI act but not to let go, adopt Watchers with a human in the loop: autonomous detection with evidence is the safest first rung, well before automatic fixes. And if you hate the idea of an agent opening pull requests on your behalf, your caution is justified — that is exactly why Grafana made Sandboxes an isolated environment with mandatory human review. In 2026 the dividing line is no longer between those who automate and those who do not: it is between those who observe their agents and those who merely trust them.