AWS adds AgentCore identity controls for production agents

TLDR: AWS ships serious security primitives for agents, OpenAI opens its networking stack, and CopilotKit raises big to own the UI layer.

Secure agent OAuth flows with Amazon Bedrock AgentCore Identity

Amazon Web Services introduced Amazon Bedrock AgentCore Identity as a standalone service that secures how AI agents authenticate to external services across Amazon Elastic Container Service (Amazon ECS), Amazon Elastic Kubernetes Service (Amazon EKS), AWS Lambda, and on premises as of 2026-05-06. The reference implementation walks through Authorization Code Grant (3 legged OAuth) on Amazon ECS with secure session binding and tightly scoped tokens.

This matters if you are moving from toy agents to production workflows that call SaaS APIs or internal services. Instead of rolling your own OAuth glue and token storage, you get a consistent pattern that plugs into existing AWS identity and access management controls. The catch: you still have to design scopes, rotation, and least privilege properly, this just gives you safer plumbing.

Expect similar identity patterns to show up in other AgentCore components so agents can traverse multiple backends without spraying long lived tokens everywhere.

Read more →

---

OpenAI publishes MRC networking protocol for AI supercomputers

OpenAI announced Multipath Reliable Connection (MRC), a new supercomputer networking protocol released through the Open Compute Project to improve resilience and throughput in large scale AI training clusters as of 2026-05-06. MRC targets the fabric between thousands of GPUs where packet loss, congestion, and link failures directly slow training.

If you work on infrastructure for big clusters, MRC is a rare peek into how OpenAI handles networking at supercomputer scale. An open spec means vendors and hyperscalers can co evolve hardware and firmware around a shared design rather than reverse engineering proprietary stacks. There are no public benchmarks yet, so treat performance claims as directional until independent evaluations land.

For smaller on premises clusters, the ideas in MRC are still useful: multipath routing, fine grained failure handling, and congestion control patterns that you can adapt even without identical hardware.

Read more →

---

CopilotKit raises $27M to standardize agent UI protocols

CopilotKit raised 27 million dollars to build infrastructure for in app AI agents that understand user actions and present interactive interfaces instead of plain chat as of 2026-05-06. The company is pushing an open source AG UI protocol that defines how AI agents connect to user interfaces and sync state.

For teams shipping user facing copilots, CopilotKit is trying to become the React for agent interfaces: one abstraction for tool calls, UI state, and user intent across frameworks. Reports say major cloud and AI framework providers have already adopted the AG UI protocol, which gives it a credible shot at becoming a de facto standard. The risk is obvious: vendor ecosystem lock in if the spec evolves behind closed doors.

If you are currently wiring agents to frontends by hand, this is worth a prototype to see whether the protocol matches your mental model of actions, context, and streaming.

Also covered by: Zamin.uz

Read more →

---

Quick Hits

• Introducing OS Level Actions in Amazon Bedrock AgentCore Browser Amazon Bedrock AgentCore Browser now exposes OS level control via the InvokeBrowser API, so agents can act on full desktop screenshots and native UI, not just the web DOM.

• SoundHound AI Introduces OASYS: The World’s First Self-Learning Orchestrated Agentic AI Platform Where AI Builds AI SoundHound AI launched OASYS, a platform that automatically creates, orchestrates, evaluates, and iterates on multiple agents across channels, pitched as an end to end lifecycle manager for agentic systems.

• Streamlining generative AI development with MLflow v3.10 on Amazon SageMaker AI Amazon SageMaker AI MLflow Apps now support MLflow 3.10, adding better observability and evaluation features for generative AI experiments so you can track, compare, and ship models with less custom glue.

• Your LLM Is Only as Good as What It Retrieves Weaviate shares a research driven overview of retrieval quality in retrieval augmented generation systems, with concrete guidance on evaluators and index design for teams who keep blaming the model instead of the retriever.

• Show HN: Airbyte Agents – context for agents across multiple data sources Airbyte launched Airbyte Agents as a unified data layer over its connector ecosystem so agents can discover data and take actions across operational systems without point to point API wiring.

• llm-echo 0.5a0 Simon Willison released llm echo 0.5a0, a fake model plugin for his LLM CLI that simply echoes input, useful for writing automated tests and for probing tools that integrate with the CLI without burning tokens.

• Secure Marketplace credentials with Production-only access Vercel now lets you mark integration resources as production only so sensitive environment variables are hidden from non production environments and require rotation if you change the setting.

• Query observability metrics using the Vercel CLI Vercel teams with Observability Plus can now query metrics directly from the vercel metrics CLI command, which is handy for coding agents that need to inspect performance or reliability on demand.

• Adding Benchmaxxer Repellant to the Open ASR Leaderboard Hugging Face updated the Open ASR Leaderboard with safeguards against benchmark gaming using private data so speech models are compared more fairly.

• [AINews] Silicon Valley gets Serious about Services Latent Space argues that the real money in AI is shifting toward durable application layer services, not raw models, tying together several recent product launches.

• How KIKO Milano scales for Black Friday Case study on how KIKO Milano used Vercel to cut build times by 75 percent, eliminate manual scaling for peak events, and move from rare releases to multiple deploys per day.