The AWS SMGS team shared a continuous quality evaluation framework built for a production-grade multi-agent sales system, centered on a nine-stage closed loop based on Amazon Bedrock AgentCore Evaluations, covering the full lifecycle from defining quality standards, CI/CD regression gates and canary monitoring to drift detection and feedback-driven optimization.
Reflection AI released its open-weight model Beam, built for agent coding and long-horizon tasks, with 501B total parameters, 23B activated parameters, and an effective context length of 1 million tokens. The weights are published on Hugging Face under the Apache 2.0 license, with BF16, FP8, and NVFP4 checkpoints available.
Why it matters: The original article details three deployment paths for Beam on AWS, along with instance selection and parameter VRAM considerations, insights transferable to planning deployments of similar open-weight large models.
Meta's coding-optimized open-weight model Muse Spark Open is now available on Amazon SageMaker JumpStart. The MoE model has 816B total parameters with 42B active per token, supports a 1M-token context window, takes text, image, document, audio, and audio-video input with text output, and is open-weight under the Muse Model License.
Microsoft Research released Magnetar, a zero-knowledge-proof virtual machine that can prove the execution of programs written in Rust, Go, or C and compiled to RISC-V, with verification of a proof taking about 12 milliseconds.
Mistral has released Timestral, its first family of time series foundation models in four sizes — 5M, 28M, 123M and 209M parameters — with the largest model ranking first among pretrained zero-shot models on the GIFT-Eval, TIME and fev-bench public benchmarks, beating the previous SOTA with about two-thirds the parameters.
Why it matters: The article gives verifiable details on the four sizes, benchmark rankings and CPU inference latency, letting readers evaluate it for time series forecasting deployments.
AWS released its Trust in Depth methodology, which lets agents operate safely without per-step human approval through four layers: Scope (limiting capabilities and access), Know (Agent Registry and independent identities), Bound (AgentCore Policy enforced at the Gateway, Guardrails, and the open-source sandbox Strands Box), and Observe (CloudWatch Omni tracking and evaluation).
A multi-agent intelligent document processing (IDP) system built on Amazon Bedrock AgentCore can automatically classify, extract, verify and learn from business documents. The solution uses Strands Agents to orchestrate 8 specialized agents, with Amazon Textract for text extraction and S3 Vectors for vector matching; new document types go through a training workflow that generates and saves extraction instructions for reuse, while known types go straight through the automated pipeline.
Google Research has publicly released Deep Review, an agent system that provides rigorous feedback and technical verification for complex technical documents. Formerly the Paper Assistant Tool (PAT), it partnered with four conferences — STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027 — to deliver pre-submission feedback on over 16,000 opted-in manuscripts.
Why it matters: The article details real-world usage of over 16,000 manuscripts across four top conferences and the four-stage agent pipeline design, which can transfer to your own document review workflows.
Amazon SageMaker JumpStart provides a catalog of pre-trained models that can be deployed via the SageMaker Python SDK, used to run repeatable performance tests, and then have their endpoints deleted.
Tamnoon built Tami on Amazon Bedrock, an agentic cloud security remediation orchestrator that can automatically investigate security findings from tools such as CNAPP and CSPM and safely orchestrate remediation.
This AWS tutorial shows how to build an AI agent system with Amazon Bedrock AgentCore that automatically turns quarterly reports, compliance documents and market analyses into client PowerPoint presentations; in internal testing the whole process took under 90 seconds.
Multiple cities across the EMEA region, along with software partners, are using NVIDIA accelerated computing to build physical AI agents and digital twins for urban planning, predictive maintenance, and public services. France's Côte-d'Or department deployed a sovereign digital twin based on Nemotron and Omniverse across more than 500 public buildings, cutting document retrieval time from weeks to minutes, lowering operating costs by 10%, with a target of 30% at scale.
Anchor Browser integrates with Amazon Bedrock AgentCore payments, letting AI agents autonomously pay for paywalled pages, APIs, and MCP servers via the x402 protocol when they encounter HTTP 402 Payment Required responses, within a configured budget.
GitHub has introduced a context-aware secret classifier that distinguishes real credentials from similar false positives via a two-stage pipeline (regex candidate generation plus fine-tuned ModernBERT classification). Compared with its previous LLM-based pipeline, weighted alert precision improved 11.5%, median post-push scan execution time dropped 86%, and service costs were cut by more than 98%.
Amazon Bedrock launched managed entitlements to simplify multi-account access to third-party AI models: subscribe once to an AWS Marketplace Private Pricing Offer in a central account, then distribute access across the entire organization.
ShipStation Global's engineering team built Deckhand, an AI IT help desk agent running on Slack atop Amazon Bedrock AgentCore. In its first three months it answered 97.5% of channel messages, with its positive feedback rate rising from 82% in April to 96% in June.
Orca Security partnered with AWS to build Discovery AI on Amazon Bedrock, converting natural language questions into Sonar queries covering 900+ cloud models and 18 operators tracked by Sonar.
The AWS team implemented kernel optimizations for Fused MoE models on Amazon SageMaker AI; the Swap-AB trick flips matrix multiplication before entering PTX assembly to fit the M=64 dimension requirement of NVIDIA Hopper's WGMMA instructions, delivering a 2–4% end-to-end throughput gain on H200, and has been upstreamed into vLLM 0.24.0 and later.
Grok 4.7 is now available on Amazon Bedrock. The author built a tool-calling agent on it and benchmarked reasoning effort levels with three tasks brute-force verified to have a single correct answer: 67 of 72 runs answered correctly; the xhigh level got 18 of 18, while the low level got 15 of 18 with output tokens about one-eighth of xhigh, and median time 18 seconds versus 106 seconds.
Why it matters: The article gives correctness rates and token costs per reasoning level across 72 verifiable runs, plus transferable evaluation and monitoring methods.
Amazon Bedrock now supports OpenAI's GPT-6 family via global cross-Region inference from Thailand, Malaysia, Singapore, Indonesia and Taiwan: GPT-6 Sol and GPT-6 Luna are available from all five AWS regions, and GPT-6 Astra from the Asia Pacific (Singapore) region.
The AWS Machine Learning Blog published a tutorial showing how to use multi-node Ray LLM DLC on Amazon EKS to deploy a single LLM across two GPU nodes, achieving prefill/decode disaggregated inference through a single OpenAI-compatible endpoint.
AWS describes two paths for deploying decision models with llama.cpp on Amazon SageMaker AI: serving OpenAI-compatible chat completions via AWS DLC, or using a SemIf (formerly OpenJev) sub-image to directly return probabilities of declarative options without decoding answer tokens.
Postman and AWS detail the production architecture of its Agent Mode, an AI-native capability covering API testing, documentation, discovery, and implementation, serving 40 million developers.
Why it matters: The piece explains Postman's concrete practices in production—controlling tool scale, using schema reads instead of piling up tools, and treating context as the bottleneck—all transferable to other agent engineering.
Ai2's AI Infrastructure team replaced its GPU cluster scheduler's priority system with one based on GPU time budgets, hierarchical fair-share and time-slicing contracts, to keep the cluster full while prioritizing high-impact research.
AWS has released a multi-tenant reference architecture based on Amazon SageMaker HyperPod and EKS, letting multiple teams share a single GPU cluster while maintaining isolation and fair resource allocation.
AWS has introduced real-time ACL enforcement for Amazon Quick and Amazon Bedrock Knowledge Bases, verifying permissions directly against authoritative data sources such as Google Drive and SharePoint at query time.
Microsoft Research Asia proposed the Harnessed Agentic RL training paradigm and open-sourced a rebuilt Agent Lightning v1.0, a roughly 3,500-line framework that lets the real agent harness used in deployment participate directly in reinforcement learning via an LLM proxy, with no need to reimplement the agent inside the training framework.
Why it matters: The article explains the design ideas, architecture components, and reproducible training results of Harnessed Agentic RL, letting readers assess its fit for existing agent training workflows.
Cornerstone OnDemand built the multi-agent system Orion AI on Amazon Bedrock and Strands Agents, cutting database incident diagnosis time from 45 minutes to 10 minutes, a 78% reduction, reducing manual database lifecycle steps from more than 10 to a single interaction, and cutting median redundant alerts by 65%.
AWS demonstrated how to build automated remediation workflows using AWS Lambda Durable Functions, Amazon EventBridge, and Amazon Bedrock, connecting to the troubleshooting results from AWS DevOps Agent.
Qlik Answers, built on Amazon Bedrock, has been generally available since February 2026 and provides sourced, verifiable natural-language question answering to more than 40,000 customers worldwide, while its Discovery Agent has surfaced over 100,000 anomalous results.
AWS demonstrated an evaluation approach for a multi-agent supply chain decision system built on the Strands Agents SDK and Amazon Bedrock AgentCore, consisting of an orchestrator agent and four specialized subagents for optimization, delivery, routing and analytics.
Amazon Quick's console and update-user API do not support directly downgrading from Admin or Author to Reader, returning a 'You cannot downgrade a user role' error.
AWS demonstrated on Amazon Bedrock Managed Knowledge Bases with LangChain a comparison of standard versus agentic retrieval: for a multi-part question containing six intents, single vector retrieval at 5 results covered only 4 sub-intents, while agentic retrieval used a planning loop to break the query into sub-queries, evaluate evidence and iterate.
Amazon Quick resources (chat agents, action connectors, knowledge bases, flows, spaces) previously had to be manually rebuilt when promoted from development to production accounts.
The aws-ai-ml skill, delivered via the Agent Toolkit for AWS, brings optimized generative AI inference on Amazon SageMaker AI to coding agents such as Kiro, Claude Code, and Codex, injecting inference optimization and benchmarking expertise.
Claude Opus 5.5 and Claude Sonnet 5.5 are now available on Amazon Bedrock in the AWS GovCloud (US) Region, and can be used with Anthropic's agentic coding tool Claude Code to support regulated development workloads, including ITAR.
Z.ai (Zhipu AI)'s GLM 5.3 is now available on Amazon Bedrock. It is a 753B-parameter MoE model aimed at coding and long-horizon agent tasks, and Z.ai reports a leading score of 84.5 on the CyberGym benchmark plus a 50% improvement over GLM 5.2 on its internal coding benchmark.
AWS launched the ability to create and manage SageMaker Spaces on HyperPod EKS clusters directly from the SageMaker Studio UI, letting data scientists launch JupyterLab or Code Editor environments in a few clicks without command-line tools.
AWS explains how to manage Amazon SageMaker HyperPod clusters through SageMaker Unified Studio, letting ML teams launch training and fine-tuning jobs from project workspaces while maintaining governance controls.
The AWS Machine Learning Blog published a tutorial demonstrating how to build Sprout, a gardening assistant with long-term memory, using the OpenClaw agent framework and Amazon Bedrock AgentCore runtime and memory.