AWS AI Updates: November 5, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. WorkSpaces for AI Agents Adds Linux and Windows 11 Cloud Desktops
Amazon WorkSpaces for AI agents now provisions Linux and Windows 11 cloud desktops for agents, in addition to the Windows Server desktops it supported before, so each agent can run on the operating system its workflow needs. On Linux, agents get full root access for developer tasks such as builds, tests and scripts; on Windows 11, they can operate the same enterprise applications employees use day to day. Both new operating systems carry the same controls as the Windows Server desktops: network isolation, IAM-based access and existing compliance boundaries. Agents reach the desktops through the service’s managed Model Context Protocol (MCP) server, and human observers can watch a session live and stop it, according to the getting-started documentation. The announcement does not name the supported Linux distributions, Regions or any pricing change. Source
2. AWS Publishes a Sample That Plugs Jev and Strands Decider 2B Into AgentCore Evaluations
AWS published a sample that connects two decision models to Amazon Bedrock AgentCore Evaluations, which ships 15 built-in evaluators plus custom LLM-as-a-judge and code-based ones: TypeSafe’s hosted Jev (requires a TypeSafe API key) and the open-source Strands Decider 2B, self-hosted on a Strands Decider server locally or on EC2 inside the customer’s VPC. Both run as code-based evaluators in an AWS Lambda function that receives OpenTelemetry spans; a spans.py module normalizes them, with trace-level evaluators sending the target turn plus brief history and session-level evaluators sending the whole conversation. Questions can be yes/no (returning the probability of yes), choice (up to 255 options, with the full distribution and a confidence) or score (2 to 10 ordered levels, scaled to [0, 1]), and each backend ships Groundedness, Helpfulness (4-level score) and GoalCompletion (3-way session choice) evaluators. They are added with evaluate.py --with-jev --jev-secret-arn <ARN> or --with-decider, and --decider-ec2 provisions the EC2 host, security groups, IAM role and Lambda VPC configuration. Before using them for release gating, AWS recommends measuring agreement with majority-vote labels from three human labelers, split into unanimous and split-vote cases, and setting per-evaluator thresholds from calibration rather than 0.5; it also suggests a cascade in which the decision model screens every turn and only low-confidence turns go to a built-in LLM evaluator, but reports no benchmark results. Source
3. AWS Details Four Context Engineering Layers in Its Open-Source Deep Insight Multi-Agent Analyst
AWS solutions architects published part two of a series on Deep Insight, an open-source multi-agent data analysis solution built on the Strands Agents SDK and Amazon Bedrock AgentCore, breaking its context engineering into four layers. For architecture, a Coordinator, Planner, Plan Reviewer and Supervisor direct tool agents (Coder, Validator, Reporter, Tracker, Auditor) that each keep their own context, share only four state keys (messages, clues, full_plan, history) and pass compressed results in a fixed CLUES_FORMAT, which keeps the Coder under 25K tokens. For prompts, each agent has an explicit output budget (Coder 1,000 to 1,500 tokens, Validator 800, Reporter 1,000) and a shared status format, and the Supervisor will not advance until every checklist item in the current section is marked done. For tools, a write_and_execute_tool writes and runs a script in one step and returns only a short summary plus the first 500 characters of output, which the authors say cuts per-script context growth by about 95%, while full results go to all_results.txt for agents to read on demand and loading skills on demand shrinks the system prompt by about 97%. For verification, Validator and Auditor agents check the work and record sources in citations.json; the authors say the example task, a sales analysis delivered as a DOCX report, would otherwise consume more than 150K tokens over about 15 minutes. A third part on harness engineering is planned. Source