#

Key takeaways

• AWS published a walkthrough for previewing web apps with AgentCore runtime Session endpoints.

• Each agent runtime session runs in an isolated microVM.

• The Session endpoint provides an authenticated URL for allowed HTTP and WebSocket traffic.

• Requests outside the allowlist, and unauthenticated requests, are refused.

• The example keeps a Vite development server running while a browser preview updates.

What AWS demonstrated

AWS published an October 30, 2026 walkthrough that shows a specific technical pattern. It uses Amazon Bedrock AgentCore runtime Session endpoints to preview web applications while a coding agent edits them. The source presents this as a published demonstration of an existing capability. It does not present it as a new general hosting service.

The example centers on a live development loop. A coding agent keeps a Vite development server running inside its own session. As files are saved, the browser preview updates. A local preview interface also places chat beside the running website. That setup lets the developer watch the app and the agent work at the same time.

How Session endpoints work

AWS says each agent runtime session operates in an isolated microVM. The Session endpoint supplies an authenticated URL. That URL forwards HTTP methods, paths, query strings, request bodies, and WebSocket connections. It only forwards traffic to explicitly allowed ports.

The sample uses port 8090 for the site and 8081 for a status route. Agent invocation uses a separate configured route. Requests to ports outside the allowlist are refused. Unauthenticated requests are also refused. These controls are part of the example's access model.

AWS also notes that Session endpoints work with microVM runtimes rather than the Instances compute type. Inbound access uses public HTTPS. The feature does not support AWS PrivateLink. The local endpoint binds to loopback, which keeps the local preview interface tied to the local machine.

Operational and governance considerations

The source highlights a few operational points that matter for anyone evaluating the pattern. Website traffic and agent activity still share a session. That means permissions must be limited appropriately. The example also records calls in CloudTrail as InvokeAgentRuntimeProxy events. That gives a trace of activity through the runtime proxy.

The walkthrough also notes a persistence limit. Edits may not survive replacement of session compute unless persistent storage is configured. That is an important design detail for any workflow that depends on saved changes. It means the preview loop is useful for live editing, but state handling still needs attention.

These details frame the feature as a controlled development environment. The source supports a view of Session endpoints as a way to connect a browser preview to an active agent session. It does not support claims about production readiness or zero security risk. It also does not support claims about universal availability.

Morocco relevance

The source reports no Morocco-specific fact. A conditional global lesson is that teams should verify access controls, session isolation, and persistence before relying on any live preview workflow.

Why this matters for developers

This walkthrough is useful because it shows how a browser preview can stay connected to a live coding session. The agent edits files, the server keeps running, and the preview reflects changes quickly. That can make iterative development easier to observe.

The example also shows the limits of the pattern. Only allowed ports are reachable. Only authenticated requests are accepted. The session is isolated in a microVM, but the website and agent still share that session. Those constraints are part of the design, not side notes.

For readers, the main lesson is simple. This is a technical demonstration of live preview inside AgentCore runtime, not a broad hosting announcement. The source supports a narrow and practical use case. It shows how a session endpoint can bridge a browser, a development server, and an editing agent in one controlled workflow.