Most AI agent discovery tools work by querying platform APIs — Salesforce Agentforce, Copilot Studio, ServiceNow. That approach has a hard ceiling: it only sees agents built on platforms that expose an API for them. Nudge Security's expanded feature, released as part of its ongoing research preview, tries to close that gap by watching the browser instead of the API. Customers toggle on browser-based discovery in a browser extension, and it picks up agents built inside tools like Cursor automations, OpenAI Agents Workflows, Zoom AI Workflows, Atlassian Rovo, Retool, Zapier Agents, and HyperAgent — platforms that, per Nudge's CTO Jaime Blasco, don't offer the same API surface for inventory. The stated problem is a blind spot, not a hypothetical one: agents built inside tools without robust public APIs are often the ones with the deepest, most persistent access to corporate systems, precisely because employees built them quickly to solve a task and moved on. Nudge frames this as a growing gap between the agents an organization thinks it has cataloged and the agents actually running with live credentials against production data. What's more interesting than the browser-hooking mechanism is the accountability layer bolted onto discovery. Once an agent is found, Nudge assigns it a technical contact — who may not be the original builder — and sends that person a 'nudge' asking them to confirm intent, justify the business case, or remediate a flagged risk. The responses populate the agent's intent field automatically, turning what would otherwise require a manual audit conversation into a passive paper trail. Every agent, whether discovered via API or browser, lands in a single inventory alongside flags for the failure modes that keep showing up in agent security research: publicly accessible endpoints, hardcoded credentials, unauthenticated connections to other systems, and orphaned agents whose creator has since left the company. This is discovery infrastructure solving a narrower problem than usual — not 'how do we find every agent on the internet' but 'how do we find every agent an employee spun up in a tool that never told IT it existed.' The unauthenticated-MCP-connection flag and the orphaned-agent flag both point at the same underlying issue seen elsewhere this year: agents outlive the accountability structures set up around them, and inventory systems are racing to catch that drift before it becomes an incident.