Field Notes

Secure RAG answers, faster incident analysis, and GPT-6 with interaction UI

Today is about rights checks in RAG, measurable effects from multi-agents, and the new GPT-6 in ChatGPT.

Key AI news

AWS adds real-time rights checks to RAG

AWS describes a two-stage architecture for Amazon Quick and Bedrock Knowledge Bases. First filters in the index. Then a real-time ACL check. That means document access lists checked directly at SharePoint, Google Drive, or Confluence. This lowers leak risk when rights get revoked or changed. Status: vendor architecture note, no independent review named. If you plan RAG with sensitive sources, you cut duplicate maintenance of permissions.1

GPT-6 rolls out with Intelligent UI in ChatGPT

OpenAI brings GPT-6 worldwide in ChatGPT. The interface delivers faster answers with visuals and interactive elements that you can use directly. Those are vendor claims. You prepare analyses right in chat, without jumping to Excel.2

Orion AI: Cornerstone shortens DB diagnosis by a lot

Cornerstone OnDemand built a multi-agent solution on Amazon Bedrock and Strands Agents. Diagnosis times fell from 45 to 10 minutes, according to the vendor. Manual lifecycle steps fell from 10+ to one interaction. Redundant alerts dropped by a median 65 percent. Three people built this in six months. If you run database operations, the case shows a path to automate handoffs and tickets.3

Highlights for your day-to-day

  • Automated remediation after AWS DevOps Agent. You route the finished RCA into a fix flow that needs approval. Durable Functions, long-running stateful workflows in Lambda, pause, wait for approval, then continue. Mutating actions run only after approval. Read-only tools run autonomously. Useful if you want to cut MTTR but keep control.4

  • Cloud-native agent control with Mecatl. An agent harness, the runtime shell that controls agents, separates at Mecatl the agent loop from tool execution and memory. Runs in Kubernetes, is centrally controllable and pausable. Good when desktop agents drop in sessions and you need governance.5

  • Train agents in the real harness: Agent Lightning v1.0. The RL framework sits as a proxy between agent and model. You train without rebuilding the runtime. The code stays slim, Kubernetes is supported natively. In one example Qwen3.5-9B rose from 41.8 to 56.4 percent Pass@1 on SWE-bench Verified with about 6,000 samples, according to the vendor.6

Tools and updates

  • Radisson ChatGPT plugin. Find, compare, book hotels right in chat. For travel and customer service flows. Prerequisite: ChatGPT with plugin access.7
  • Mistral Large 4 announced. New model for complex text tasks. Check access and license before you move workloads.8
  • n8n + Vizard AI. Cut long videos into clips automatically, post only clips over threshold, log to Sheets. Prerequisite: n8n workflow and Vizard account.9
  • nanoMuse Paper. Open personal agent for your own devices and accounts. For teams that want to evaluate a self-hostable agent concept. Requires willingness to experiment.10

Try it in five minutes

1) Open ChatGPT and start a new chat with GPT-6. Ask for a visual, interactive summary of this sample data: Product A 1.2 million revenue Q3, churn 3%; Product B 0.8 million, churn 6%; Product C 0.5 million, churn 4%. Say: "Create a visual overview and an interactive table for exploration."2 2) Add a short audience, for example "for sales leadership", and ask for two scenarios with concrete actions at churn >5 percent.2 3) Ask for export options: "Which elements can I take as an image or table?" Test a click or filter if offered.2

Expected result: An answer with visuals and controllable elements plus two action lists. Check: You can use at least one element interactively or take it as a graphic or table. If you get text only, Intelligent UI is likely not active for you yet.

A good find

AWS shows how to assess agents beyond "hours saved". Four value dimensions matter: time, exceptions including costly fixes, decision quality with consistent policy, and maintenance economics with frequent process changes. What decides is realization. Who turns the effect into lower spend or higher yield. Next step for you: Pick a process with rework, estimate fix costs with a rework multiplier, and assign a responsible person for realization.11

Sources

  1. Rethinking access control for RAG with Amazon Quick and Amazon Bedrock (aws.amazon.com)
  2. GPT-6 and Intelligent UI for everyone (openai.com)
  3. How Cornerstone OnDemand cut database diagnosis by 78% with Amazon Bedrock (aws.amazon.com)
  4. Automate remediation post AWS DevOps Agent investigation (aws.amazon.com)
  5. Can a Cloud-Native Harness Make Agents Reliable Beyond the Desktop? (latent.space)
  6. Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real Harnesses (microsoft.com)
  7. Radisson Hotel Group brings hotel discovery into ChatGPT (openai.com)
  8. Introducing Mistral Large 4 (mistral.ai)
  9. @n8n_io: Long video in, short clips out ♻️ Vizard AI does the clipping. Anything below your viral score threshold never gets posted. The rest go out and get logged to Sheets. Check it out: (x.com)
  10. Paper: nanoMuse: An Open-Source Personal Agent for Every Device You Own (huggingface.co)
  11. Beyond hours saved: Building the business case for agentic automation (aws.amazon.com)