Aidan Sochowski describes how Runlayer found more than 300 agents running daily across a company of roughly 50 to 60 people. The team used a gatewayAn AI agent gateway is a shared control layer that mediates agent access to models, tools, data and observability services. and centralized session and tool-call logs to map agent activity and identify systems exposed to risky actions.
Sochowski illustrates prompt injectionPrompt injection occurs when untrusted content attempts to redirect an AI system's behaviour beyond the authority of that content. through incoming support emails and the possibility of a code-review agent being persuaded to approve a sensitive change. He describes assessing tool calls in the context of a whole session, comparing behavior with an agent's intended task and narrowing access to data and shared systemsAn agent permission boundary defines and enforces which resources and actions an automated agent is allowed to use..
The talk also covers monitoring unreliable MCP servers and repeated errors that increase token costs. Sochowski discusses rate limits, spending budgets and model choices as ways to manage costs while retaining visibility into agent behavior.
Watch on YouTube
