Bedrock AgentCore added authenticated session endpoints so a coding agent's web app stays live in its own sandbox and edits show up in the browser within seconds
AspireAI community · Amazon Web Services Machine Learning Blog
The fastest way to review a coding agent's work is to open the app it built, use it, ask for a change and check the result - but that loop only works when the app keeps running where the agent edits it, because copying it out and deploying elsewhere makes every change wait on a deployment. Amazon Bedrock AgentCore runtime runs each agent session in its own isolated microVM, and the new Session endpoints give each session an authenticated URL that forwards each request's method, path, query string and body, plus WebSocket connections, to the ports you allow, so a browser can open an app running inside the agent's session. The post walks through deploying a coding agent that keeps a Vite site running in its session and opening it in a split preview, with the agent's steps streaming back while Vite hot-reload pushes each saved change to the browser without a refresh. Every request is authenticated - SigV4 for IAM-authenticated runtimes, a bearer token where an inbound JWT authorizer is configured - so knowing the URL grants nothing, and the URL dies with the session. The limitation is explicit in the post: session endpoints are for building, testing and reviewing, and serving to end users means deploying to a hosting service such as Amplify Hosting.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.