A single Fortune 500 company will deploy thousands of AI agents in the next 18 months. Most of them have no idea what they already run. That visibility gap is where AWS is placing its bet with a quiet but comprehensive infrastructure play inside Amazon Bedrock AgentCore.
The company launched three interconnected capabilities this week that, taken together, solve the three things preventing enterprise AI agents from graduating out of pilot programs: discovery and governance, human oversight, and stateful interactivity. None of these are glamorous. All of them are essential.
Agent Registry, now in preview, gives platform teams the one thing they lack when agent sprawl hits: a single pane of glass. The registry indexes every agent, tool, MCP server, and agent skill across your entire estate—AWS, other cloud providers, on-premises. It captures metadata: who published it, what protocols it implements, how to invoke it, and whether it's compliant. You can register records manually through the console or API, or point the registry at an MCP or A2A endpoint and have it pull details automatically. The governance layer includes approval workflows for publishing and fine-grained control over who can discover and consume what. For organizations that have watched their agent portfolio grow organically across teams, this is the audit trail they need to sleep at night.
BrowserLiveView solves a different problem: trust. When an AI agent navigates a website on a user's behalf—filling forms, pulling records, executing transactions—the user has no visibility into what's happening. BrowserLiveView embeds a real-time video feed of the agent's browsing session directly into a React application using just three lines of JSX. The component uses Amazon DCV protocol to render the stream, and your server only needs to generate a presigned URL—no streaming infrastructure to build. For regulated workflows where a supervisor must observe agent behavior in real time and intervene if needed, this is the interface layer that makes human-in-the-loop practical. It also generates visual audit evidence for compliance teams who need to prove what the agent did.
The third piece is Stateful MCP, which completes the bidirectional Model Context Protocol implementation on AgentCore Runtime. The original MCP server support was stateless: each HTTP request was independent, which works fine for simple tool calls but breaks down when a workflow needs to pause mid-execution for clarification. Stateful mode provisions a dedicated microVM per user session, enabling three capabilities the MCP specification defines but stateless servers can't deliver: elicitation (requesting user input mid-execution), sampling (invoking LLM generation from the client), and progress notification (streaming real-time updates during long-running operations). This transforms one-way tool execution into actual conversations between MCP servers and clients.
AgentCore positions itself as the substrate that ties these pieces together: open to any model, any framework, any enterprise architecture. AWS is not trying to own the agent logic itself—it wants to own the control plane underneath. For platform teams standardizing on AWS infrastructure, this creates a single namespace for all agent operations regardless of where agents actually run.
The competitive picture matters here. Google's Vertex AI Agent Builder, Microsoft's Azure AI Agents, and a growing field of point solutions from LangChain, CrewAI, and others are chasing the same enterprise market. AWS's advantage is not technical novelty—it's distribution. The company already holds the cloud relationships and compliance certifications that financial services, healthcare, and government contractors require. If those customers standardize agent infrastructure on the platform where they already run their workloads, AWS wins by default.
The deeper point: enterprise AI agent adoption is not a model problem. The frontier models are good enough today. It's an operational problem. You need to know what agents you have, you need visibility into what they're doing, and you need humans in the loop for consequential decisions. AWS AgentCore now addresses all three axes simultaneously. Platform teams can finally point to a single control plane and say: this is how we discover, monitor, and govern every agent in production.