MCP is emerging as the bridge connecting AI agents to infrastructure data, evidenced by Datadog shipping an MCP server that exposes observability dashboards to AI agents for automated responses. In parallel, Qualys flagged security risks as 53% of MCP servers rely on static secrets. These developments signal a shift in how AI systems access telemetry data, moving beyond traditional metric pipelines.

Two architectural approaches are emerging for MCP observability. The first, exemplified by Datadog, wraps existing platforms to provide pre-processed data. The second builds MCP-native observability directly at the kernel level, as demonstrated by tracing a vLLM regression where an AI agent identified logprobs computation blocking the decode loop—something invisible in aggregate metrics. This direct-to-source approach enables precise root-cause analysis.

Security concerns around MCP servers remain valid, particularly for those accessing sensitive infrastructure data. The future trajectory suggests expanding eBPF-based observability beyond GPUs to networks, security, and cost monitoring. Rather than wrapping existing tools, the emerging pattern gives AI agents direct access to raw telemetry, allowing them to determine what data matters most.