AWS

Amazon Bedrock AgentCore has launched a Web Search API that lets agents fetch current web results with title, URL, publication date and query-relevant extracts in a single call, served entirely from an Amazon-operated index rather than a third-party search engine. It supports domain include/exclude filters and publication-date ranges, and can be paired with LangGraph to build production workflows such as daily dependency watchers.

AWS introduced agent evaluation capabilities in Bedrock AgentCore, offering built-in, DeepEval and AutoEval evaluators retrievable through a single ListEvaluators API call. The post argues that surface signals like latency and user feedback cannot reveal task completion, long-session topic drift, PII leakage or skipped steps, so teams should define success criteria and trace-level evidence before choosing evaluators.

TCS built an agentic test orchestrator on Amazon Bedrock AgentCore to automate the chain of BFSI testing work, from requirement interpretation and scenario design to automation authoring, environment setup, and failure triage. The design emphasizes project-specific grounding, combined API and browser UI coverage, and tenant-level data isolation, aiming to raise testing throughput while preserving financial-services controls.

AWS released a centralized observability data lake for Amazon Bedrock that captures per-invocation prompts, responses, token usage, and model identity, enabling per-call cost derivation. The solution ships with CloudFormation templates, Lambda code, and SQL queries on GitHub, helping platform and ML engineers debug spikes, spot wasteful prompts, and attribute usage across teams.

An AWS blog post explains how agents, skills, and connected data sources fit together in AWS Quick: agents interpret natural-language requests and orchestrate connectors, while skills encode repeatable tasks such as weekly status updates, release notes, and ticket creation into reusable, team-shareable instruction units. It also covers combining the official skill catalog with custom agents, but remains a vendor-bound conceptual and how-to guide without new models or quantitative results.

Amazon Bedrock Managed Knowledge Base now supports single-shot RAG and retrieval-only requests alongside its existing agentic retrieval. Developers can match retrieval effort to each query, switching between fast single-pass search and deeper multi-step search without building separate systems.

AWS released a seven-layer composable reference architecture for enterprise agentic AI on Amazon Bedrock AgentCore, with an open-source sample implementation. It targets the common gap between prototypes and production by front-loading routing, guardrails, observability and security as reusable layers rather than afterthoughts.

OpenAI announced that ChatGPT Enterprise is now available through AWS Marketplace, letting organizations discover, buy, and manage the product via their existing AWS purchasing and billing relationship, with eligible spend counting toward AWS commitments. The listing lowers procurement friction and marks a channel expansion for OpenAI's enterprise business.

LangChain announced general availability of LangSmith Bring Your Own Cloud (BYOC) on AWS, letting enterprises run a managed LangSmith deployment inside their own AWS account and VPC. Traces, datasets, experiments, prompts, agent deployments and sandbox data stay within the customer's cloud boundary, while LangChain handles monitoring, upgrades, scaling and cluster lifecycle management.

AWS engineers at the Linux Plumbers Conference described how their EC2 kernel team filters out roughly 70% of kernel CVEs with scripts, then hands the rest to AI agents that produce triage tables using the source tree, configs and threat models, with human review still required. The team also advocates moving to newer LTS kernels to avoid accumulating unbackported fixes on old branches.