Skip to content
MustHave.ai
MustHave Hoodie Seal
THE MUSTHAVE NEWSROOM

AI news

What’s new. What matters. What do you think?

Tuesday, September 15, 2026

Sep 15

Ninth Wave Banking AI Uses Seven Agents but Keeps Scoring Deterministic

Ninth Wave banking AI uses an orchestrator and seven specialist agents for onboarding analysis, while code keeps the readiness score deterministic.

GitHub Copilot Custom Properties Now Suggest Governance Metadata

GitHub Copilot custom properties can now suggest allowed metadata values for repositories. The preview can speed setup, but policy review still matters.

NVIDIA Says DSX Raised AI Token Throughput 24% Without More Power

NVIDIA reports 24% more B200 token throughput within the same power budget. A new alliance with Google and Emerald AI now targets flexible-grid standards.

Claude API Adds Background Conversation Compaction for Long-Running Agents

Claude API compaction replaces old message history with a signed summary block while preserving a recent tail. This guide maps the new failure modes.

Google Launches Gemini 3.8 Live With Extended Thinking for Voice Agents

Gemini 3.8 Live adds a production audio model and an extended-thinking variant. Here are the migration changes, state signals, limits and prices.

Prompt, RAG, Fine-Tune or Train? A Practical AI Customization Guide

AWS published an eight-step model customization spectrum. This vendor-neutral guide maps knowledge, behavior, latency and cost failures to the right step.

Abnormal AI Email Security Uses Code-Executing Agents for Hard Cases

Abnormal AI describes a three-tier email security pipeline where only the hardest cases reach code-executing agents in ephemeral sandboxes.

AWS AgentCore Adds a Managed OAuth Consent Portal for AI Agents

AWS AgentCore OAuth consent now manages browser authorization and session binding for user-authorized tools. The callback map still demands careful setup.

GitHub Copilot Auto Adds Efficiency, Balance and Intelligence Tiers

GitHub Copilot Auto can now optimize each prompt for efficiency, balance or intelligence. Here is how to test cost per accepted coding task.

Monday, September 14, 2026

Sep 14

Cohere North Small Translate Has 218B Parameters, but Its Open Weights Are Non-Commercial

Cohere North Small Translate activates 25B of 218B parameters across 50 languages, with gated non-commercial weights and a high-end GPU floor.