AWS Bedrock AgentCore expands: OpenSearch Agent Skills, custom code evaluators, Nova 2 moderation

This is AWS's most coherent agent-infra week to date. OpenSearch Agent Skills targets a real pain point — search/analytics teams hopping between tools to wire up domain expertise — by shipping curated, open, composable skills that drop into agentic IDEs like Cursor, Claude Code, or the new AgentCore runtime. Custom code-based evaluators close the loop on production agent QA: developers can write a Lambda, register it as an evaluator, and run it both on-demand and online alongside built-in evaluators for grounded fact-checking and PII detection.

The Nova 2 Lite moderation prompting guide is more strategic than it appears — by anchoring on AILuminate (MLCommons), AWS is positioning Nova as the default moderation primitive on Bedrock against open-weight alternatives. Customers can swap taxonomies while keeping the structured prompt scaffolding.

The Q1 2026 token-volume datapoint (greater than all prior years combined) is the headline number for AWS's AI thesis. With OpenAI now on Bedrock post-Microsoft restructuring and Anthropic flowing strong adoption (Ramp index #1), Bedrock is on track to be the most multi-model agent runtime in market. SageMaker Studio Flexible Training Plans adding GPU reservations (cost cuts up to 65% vs On-Demand) and Trainium/EFA Kubernetes DRA support round out a stack-wide week.

Watch next: AgentCore's identity primitives and how AWS positions vs Vertex AI Agent Builder post-I/O. The Bitnami images removal from ECR Public (June 10) is a side-quest customers must action on now.