📰 AI Daily Recap

Friday was a runtime day. The platforms an AI agent actually lives in kept hardening: Microsoft shipped a serverless agents runtime in Azure Functions at Build 2026 (agents defined in .agent.md files with YAML triggers, MCP access, and sandboxed execution), Mantyx pitched a batteries-included managed agent runtime, and Anthropic quietly paused token-based billing for the Claude Agent SDK — a reminder that the commercial model for agent infrastructure is still being figured out in real time.

Underneath the runtimes, the connective tissue stayed in focus. Sean Lynch (via Simon Willison) made the sharp point that MCP's real edge over plain skills/CLI is isolating the auth flow outside the agent's context window — a framing that lines up with two security stories landing the same day: BleepingComputer arguing every AI agent is an identity most orgs don't govern, and Microsoft positioning Windows (with a new Execution Context model) as the trustworthy OS for autonomous agents. The plumbing is maturing faster than the governance around it.

On the model front, the open-weights story finally read like a frontier story: Latent Space's AINews flagged GLM-5.2 passing everyone's vibe check, with Z.ai forecasting an open Fable-class model by December. And on evals, a recurring theme — simplicity wins — surfaced in a harness claiming SOTA on SWE-Pro, Terminal-Bench 2, and SWE-Verified across 21 models with a deliberately simple agent.