AWS AI Updates: October 8, 2026

1. Claude Haiku 5.5 Launches on Amazon Bedrock, Claude Platform on AWS, and GovCloud

AWS. Anthropic’s Claude Haiku 5.5 is now available on Amazon Bedrock and Claude Platform on AWS, with AWS citing Anthropic’s figure of roughly 75% lower cost than Claude Haiku 4.5 for most tasks. It is the first Haiku with effort controls, supports high-resolution images, and is positioned as a fast subagent for coding, tool use, and computer use beneath a planner such as Claude Opus 5.5. On Bedrock it is served through US, EU, AU, and JP geo cross-Region inference profiles plus a global profile (model ID global.anthropic.claude-haiku-5-5) via InvokeModel, Converse, or the Anthropic Messages API on bedrock-runtime; AWS notes responses may include a thinking block before the text block. A same-day launch also brings it to AWS GovCloud (US) on both bedrock-runtime and bedrock-mantle endpoints, while Claude Platform on AWS access is limited to North America. Source

2. Amazon Quick and Bedrock Knowledge Bases Verify Document Permissions Live at Query Time

AWS. AWS detailed a two-stage access control design for RAG in Amazon Quick and Amazon Bedrock Knowledge Bases: a semantic search first filters candidates using ACLs cached in the index, then each candidate is checked against the source system (for example, Google Drive APIs through service-account impersonation) before any passage reaches the LLM. The goal is to close the gap left by sync-based ACL replication, where revoked access persists until the next connector sync and sources like Confluence emit no event on group membership changes. AWS says revocations are reflected in answers “within moments,” and cites Mondelez International rolling Quick out to more than 35,000 employees on that basis. Source

3. Sample Turns AWS DevOps Agent Findings Into Approval-Gated Fixes With Lambda Durable Functions

AWS. A new CDK sample takes AWS DevOps Agent’s investigation events from EventBridge and runs them through a Lambda Durable Function that loops with Amazon Bedrock to propose fixes using only an allowlist of purpose-built Lambda tools. Read-only tools run on their own, while mutating actions (such as raising a Lambda timeout or editing an IAM policy statement) checkpoint and suspend the workflow, using no compute, until a human sends an approve or reject callback; the callback accepts arbitrary JSON so reviewers can pass parameter overrides. The repo also ships Kiro rules so deployment and the simulated Lambda timeout incident can be driven through the AWS MCP Server. Source

4. Cornerstone’s Strands-Based Multi-Agent System Cuts SQL Server Diagnosis From 45 to 10 Minutes

AWS. Cornerstone OnDemand built Orion AI, a hub-and-spoke multi-agent system on Amazon Bedrock and the open-source Strands Agents framework, with a meta-orchestrator delegating to domain agents for monitoring, database diagnostics, lifecycle operations, analytics, and support. Three agents split each database investigation (blocking chains, wait types, long-running queries), cutting diagnosis from about 45 to 10 minutes; 10-plus-step lifecycle tasks became a single interaction, and redundant alerts dropped by a median 65%. A three-person team shipped it in six months, and the main design lesson offered is to split agents by domain rather than by task complexity to keep each tool context narrow. Source

5. Qlik Details Layered Agent Architecture Behind Qlik Answers on Bedrock

AWS. Qlik described how Qlik Answers, generally available since February 2026, separates a lightweight routing layer, an answer layer with fast and decomposition paths, a shared swarm runtime for specialist agents, and OpenSearch-based retrieval, all calling Amazon Bedrock through Qlik’s own LLM gateway with Bedrock Guardrails applied to every request and response. The design targets data residency across 11 Regions without maintaining separate builds, and Qlik plans token capacity 3 to 6 months ahead of launches. Its Discovery Agent has surfaced more than 100,000 anomalies for customers since launch. Source

6. Reference Build Adds a Nova 2.5 Sonic Voice Concierge to an Airline App via AgentCore

AWS. A new CDK-deployed sample builds a speech-to-speech airline concierge with Strands Agents and Amazon Nova 2.5 Sonic, hosted on AgentCore Runtime over WebSockets with per-session microVM isolation. Backend Lambda and API Gateway endpoints for itineraries, seat maps, meal preferences, flight status, and loyalty are exposed as MCP tools through AgentCore Gateway, policy questions are grounded in Bedrock Knowledge Bases, and the agent can hand off to a live agent with a reference number and estimated wait time. The stack spans nine CDK stacks using DynamoDB, Cognito, SES, and an Amplify-hosted React front end, with synthetic data for adaptation. Source