Bedrock AgentCore's web search API is GA — $7 per 1,000 queries
One InvokeWebSearch call drops a ranked web index into any agent loop. Plus: DeepMind verifies Signal, NVIDIA open-sources its Kubernetes stack, and more.

Copy markdown
Any agent gets web search in one $7/1k call
Amazon Bedrock AgentCore's Web Search API is GA: the new InvokeWebSearch call returns ranked results from Amazon's own web index — text extract, title, URL and publish time — with domain and date filters, no gateway and nothing to provision. It's $7 per 1,000 queries from the AWS SDKs or CLI, IAM-authed and live in six regions, so you can bolt real-time search onto any custom agent loop or data pipeline today.
DeepMind formally verifies the Signal protocol
AlphaProof Code, a Gemini-Carbon-powered agent framework, produced a machine-checked Lean proof that Signal's compiled cryptographic code is correct — in four weeks, work that used to take expert teams years, with humans reviewing only ~1% of it. The verification artifacts and full paper are open-sourced: a concrete template for proving your own security-critical code, not just its design.
NVIDIA open-sources its GPU Kubernetes stack
NVIDIA Kubernetes Engine (NVKE) and Kubernetes Recipes (NVKR) hit public preview, Apache-2.0 on GitHub. NVKE runs a managed-style K8s service on your own GPU hardware with per-cluster control planes; NVKR builds version-pinned, GPU-validated node and container images plus Helm charts. A local quickstart runs on a single machine with no GPU or NVIDIA account needed — a real on-ramp for standing up self-hosted inference clusters.
Cloudflare's Markdown for Agents goes GA
Cloudflare's HTML-to-Markdown conversion for agents is now GA with real observability: an Analytics dashboard and GraphQL API surface conversion outcomes, estimated token savings, top converted pages and failure reasons, plus a Playground that diffs Markdown against HTML per URL. If you feed web pages to models, you can finally measure how much context — and cost — you're cutting.
Nemotron 3.5 Super VL is one-click on SageMaker
NVIDIA's 120B-parameter vision-language model, with up to 1M-token context, is now deployable from Amazon SageMaker JumpStart via the Studio UI or Python SDK — a large VLM you can stand up inside your own AWS account for document, image and long-context workloads without wiring up serving yourself.