Live
Open‑Weight AI Models and Ansible‑Driven HPC Management: New Paths for Telco and Enterprise OpsBrowser Run switches to Chrome 152 – what engineers need to knowEvaluating GPT-6.1 Terra: Low-Cost Speed Gains vs. Accuracy Loss for LLM PipelinesOpenTelemetry Collector Powers Native Tracing and Enriched Kubernetes Events in OpenChoreoHow Gemini Video Courses for Finance and Legal Impact Your Cloud ArchitectureClaude Code session cap removed for Ultra plan – operational impact for AI engineersD1 result size limit forces query redesign for sub‑50 MB responsesBackdoored Cursor VS Code Extension Exposes Thousands of DevelopersOpen‑Weight AI Models and Ansible‑Driven HPC Management: New Paths for Telco and Enterprise OpsBrowser Run switches to Chrome 152 – what engineers need to knowEvaluating GPT-6.1 Terra: Low-Cost Speed Gains vs. Accuracy Loss for LLM PipelinesOpenTelemetry Collector Powers Native Tracing and Enriched Kubernetes Events in OpenChoreoHow Gemini Video Courses for Finance and Legal Impact Your Cloud ArchitectureClaude Code session cap removed for Ultra plan – operational impact for AI engineersD1 result size limit forces query redesign for sub‑50 MB responsesBackdoored Cursor VS Code Extension Exposes Thousands of Developers
AWS

Live‑preview web apps from a coding agent using AgentCore session endpoints

AI SummaryPowered by AI

Amazon Bedrock AgentCore now offers session endpoints that expose a running agent’s microVM as an authenticated URL, letting you view and interact with a web app in real time. This removes the deployment step from the review loop, speeding up iteration for AI engineers, platform teams, and security operators.

AgentCore session endpoints let a coding agent keep a Vite‑based web site running inside its isolated microVM and expose it through a signed URL. Practitioners can open that URL in a browser, see file changes instantly, and send WebSocket traffic without redeploying the app, which shortens the feedback cycle for prototype development.

How Session Endpoints Work

When you create a runtime with CreateAgentRuntime, you specify an allowedPorts list – up to five entries, each a single port or a range inside 8000–12000, excluding the reserved platform ports. The runtime does not open a listener automatically; your agent must bind a server to the chosen port on 0.0.0.0.

The InvokeAgentRuntime call returns a header named X-Amzn-Bedrock-AgentCore-Runtime-Session-Endpoint-Url. SDKs surface this as sessionEndpointUrl. The URL contains a host prefix that maps the external request port (always 443) to the internal port you allowed, for example https://8090--.runtime.bedrock-agentcore..on.aws forwards to port 8090 inside the session.

Every request – HTTP, HTTP/2, or WebSocket – must be authenticated. For IAM‑based runtimes, each request carries a SigV4 signature and requires the bedrock-agentcore:InvokeAgentRuntimeProxy permission on both the runtime and its endpoint. For runtimes using an inbound JWT authorizer, a bearer token is required instead. Knowing the URL alone does not grant access.

Because browsers cannot add SigV4 signatures or JWTs to outbound requests automatically, the Bedrock AgentCore SDK runs a local proxy (e.g., on http://localhost:8088) that injects the required credentials before forwarding traffic to the session endpoint.

Architectural and Operational Implications

  • Shorter iteration loops: Developers can edit code, watch the live page update, and request further changes without waiting for a separate deployment pipeline.
  • Isolated testing environment: Each session runs in its own microVM, limiting cross‑session interference and simplifying cleanup after a review.
  • Port management: Teams must ensure the agent’s web server listens on an allowed port and binds to all interfaces; otherwise the endpoint will reject connections.
  • Local proxy requirement: Production users cannot connect directly; a trusted local process is needed to sign requests, which adds a small operational component to the workflow.

Security Considerations

The endpoint enforces per‑request authentication, so accidental exposure of the URL does not compromise the service. However, the following points merit attention:

  • Credentials used by the local proxy must be protected, as they can sign any request to the session.
  • Allowed ports should be limited to the minimum required; broader ranges increase the attack surface within the microVM.
  • Because the session URL is short‑lived, it should be discarded after the review to avoid reuse.
  • When using IAM, the bedrock-agentcore:InvokeAgentRuntimeProxy permission should be scoped tightly to the specific runtime and endpoint resources.

Related CloudNinjas coverage: AWS.

What This Means For Practitioners

AgentCore session endpoints provide a practical way to preview agent‑generated web apps without a full CI/CD push. Teams should adopt the pattern for rapid prototyping, ensure port configurations match the agent’s server, and treat the local credential‑injecting proxy as a trusted component. Monitoring the lifetime of session URLs and restricting IAM permissions will keep the convenience from becoming a security liability.

Originally published atAWS Machine Learning Blog