AI and developer tool updates (2026-05-06) #10

Today's Letter

1. Google, event-driven webhooks added to Gemini API
2. AWS, AgentCore Optimization preview announced
3. Amazon SageMaker AI, agent-guided model customization workflow added

Google, event-driven webhooks added to Gemini API

• Google introduced event-driven webhooks for long-running jobs in the Gemini API
• The change shifts completion handling from repeated client polling to asynchronous HTTP callbacks
• Target workloads include Deep Research, long video generation, and large Batch API jobs that can run beyond an interactive request window
• Webhook delivery is at-least-once, so receivers must handle duplicate events safely
• Failed deliveries are retried automatically for up to 24 hours
• The update reduces polling overhead and latency for Gemini API applications that manage asynchronous agentic workflows

Source: blog.google
More: news.google.com

AWS, AgentCore Optimization preview announced

• AWS introduced AgentCore Optimization in preview for Amazon Bedrock AgentCore
• The feature analyzes production traces and evaluation outputs to recommend system prompt or tool-description changes
• Recommendations use CloudWatch Log group traces and a selected built-in or custom evaluator as the reward signal
• Validation is split into offline batch evaluation and online A/B testing on production traffic
• A/B tests run through AgentCore Gateway and report confidence intervals and statistical significance
• Runtime configuration is managed as immutable bundles tied to runtime ARNs, separating model ID, system prompt, and tool descriptions by version
• AgentCore Observability links model calls, tool calls, reasoning steps, and evaluator scores through OpenTelemetry-compatible traces

Source: aws.amazon.com
More: news.google.com

Amazon SageMaker AI, agent-guided model customization workflow added

• AWS introduced an agent-guided model customization workflow in Amazon SageMaker AI
• Developers describe a use case in natural language, and an AI coding agent helps with data preparation, technique selection, evaluation, and deployment
• The workflow is powered by nine modular skills covering use-case definition, planning, dataset validation, transformation, fine-tuning, evaluation, and deployment
• Supported fine-tuning techniques named in the post are SFT, DPO, and RLVR
• Evaluation can include LLM-as-a-judge metrics, and deployment targets include SageMaker AI endpoints or Amazon Bedrock
• SageMaker AI Studio JupyterLab has Kiro preconfigured by default, while ACP-compatible agents such as Claude Code can also use the same skill integration
• Requirements include SageMaker AI Distribution image 4.1 or later, execution-role permissions, an S3 bucket, and a configured SageMaker domain

Source: aws.amazon.com
More: news.google.com

Jocoletter curates AI, software, and product trends for developers and builders.

#AWS #Google

Subscribe to Jocoletter