Observability startup Grafana Labs Inc. said today it’s trying to shine a light on the “black box” inner workings of artificial intelligence models with the launch of new capabilities that will better enable companies to trust and control them in production.

The announcement came during Grafana’s annual user conference GrafanaCON 2026 in Barcelona, where it also revealed it’s creating a dedicated “AI organization” that will be led by its new Director of AI Mat Ryer.

Grafana said vast numbers of enterprises are racing to integrate AI agents into their workflows with a view to enhancing automation, but doing so presents numerous challenges in terms of observability. First and foremost is the fact that AI applications behave very differently compared with the traditional software that helped Grafana establish itself as a major observability player.

Because they work so differently, existing monitoring tools struggle to provide much insight into agents, making them difficult to debug. In addition, developers often find themselves constantly switching between coding environments like GitHub Copilot and Cursor and the observability tools they’re using to keep tabs on their code in production.

The startup’s solution is to treat agent sessions and large language model conversations as standard telemetry signals, similar to the logs, metrics and traces associated with traditional applications. The idea is that users will be able to monitor the performance of agents in the context of their broader information technology infrastructure.

The biggest new update is a new AI observability capability that’s launching as a public preview in Grafana Cloud today. It’s tasked with observing the behavior of AI agents, including their inputs, outputs and execution flows, so that these can be continuously monitored for low-quality responses, policy violations and other anomalous activity. According to Grafana, it’s able to surface risks such as data exposure or leaked credentials much faster than existing observability tools not designed for AI agents.

Developers will likely appreciate the new Grafana Cloud CLI, or GCX, command line tool, which is a new “agentic interface” that’s designed to live where they work. So rather than having to jump from their integrated development environment to a Grafana dashboard, they can use GCX to invoke the Grafana Assistant directly within environments like Claude Code or GitHub Copilot. The idea is to create a “continuous feedback cycle,” where developers can constantly watch their AI agents in production, correlate alerts and receive suggested code fixes in real time.

On the back end, Grafana says, it’s overhauling its core observability engine to better deal with the massive volumes of data generated by AI systems. The main focus here is Grafana’s log aggregation tool, which has been updated to Grafana Loki Evolution. Loki has been rebuilt from scratch on an Apache Kafka-based architecture and comes with a new query planner tool that accelerates performance by ten-times on aggregated queries while scanning 20 times less data.

Grafana is also making its AI-powered Grafana Assistant tool available to more users. Previously limited to Grafana Cloud only, the AI sidekick is now being integrated with on-premises Grafana Enterprise deployments to cater to customers with strict data privacy requirements. It also gains new features such as an “assistant workspace” for full-screen interactions and an “assistant API” that can extract Grafana’s insights into third-party tools.

Finally, in a nod to the open-source community, Grafana has published a new benchmarking tool called o11y-bench. It’s meant to measure how well AI agents perform various real-world tasks, such as fixing an app’s broken dashboard or investigating an outage, against a live Grafana stack. According to the company, it’s meant to be a standardized “IQ test” that measures the effectiveness of AI observability agents.

Although it went unsaid, today’s updates illustrate Grafana’s long-term vision. If AI agents are going to start running enterprises at scale, they’ll need a supervisor. By forming a dedicated AI unit, tasked with “coordinating work across AI observability, assistant experiences and agent-driven workflows,” Grafana is determined to take on that role before someone else steps up.

Image: SiliconANGLE/Microsoft Designer