Runtime monitoring design

What it is for

You cannot respond to what you cannot see. AG8 designs the telemetry and detections that show, in production, when an agent starts behaving outside its job.

What we look at

• What each agent, tool call and model call should emit, and where it goes.
• Detections for scope elevation, unexpected egress, writes to memory, newly connected tools or servers, and sudden changes in volume or cost.
• How alerts reach the people who can act, and what they need in front of them.

Standards we build on

OpenTelemetry’s semantic conventions for generative AI, which cover spans, metrics and events for GenAI clients and MCP, now live in their own repository^⁠[1] and are still marked as in development,^⁠[2] so the schema we design records the convention version it follows. The MCP security best practices recommend logging scope elevation events and all stdio transport use.^⁠[3] Managed platforms are adding this too: in June 2026 Google’s agent platform turned OpenTelemetry tracing on by default for newly deployed agents built with its Agent Development Kit.^⁠[4]

What you receive

A telemetry schema, a set of detections, and SIEM content your security operations team can load.