The protocol's biggest overhaul ships a stateless request/response core (no session IDs), MCP Apps and Tasks as versioned official extensions, and Tunnels in research preview for private-network access without exposing servers publicly.
Why it matters: MCP Tunnels mean enterprise teams can connect internal tools to Claude agents without opening inbound firewall rules — a key unblock for organizations that couldn't expose internal systems to the public internet.
xAI's Grok 4.5 is now rolling out in GitHub Copilot with a 500,000-token context window, image input support, and three reasoning-effort modes built for fast agentic coding and multi-step workflows.
Why it matters: More frontier model choice inside Copilot broadens options for engineering teams — Grok's agentic strengths now sit alongside Claude Code without requiring a separate tool or subscription.
Microsoft's first in-house cybersecurity model (trained on 100T daily signals) powers a three-agent red/blue/green platform entering public preview August 3, claiming 96% on CyberGym benchmark at 50% of its current frontier-model stack cost.
Why it matters: Enterprise buyers are paying for purpose-built vertical agents — a validation that agent products positioned by use case command premium pricing and dedicated procurement budgets.