Sep 21, 2026

Muse reaches the Mac, Google's agent kit hits 1.0, and what good approvals look like

A personal AI agent that acts inside your Mac apps, a mature agent framework for Kotlin and Android, and a practical look at the approval design both point to.

1. 01Tools

   Meta's Muse agent arrives on the Mac ↗

   TechCrunch reports Meta released Muse for Mac on 17 September. It can work with your files, messages, calendar, notes and mail inside their native apps. Meta says access to what Muse can touch on your Mac is opt-in, and that it will always ask for your approval before sensitive actions. The article does not say whether it runs locally or in the cloud, or give pricing; its earlier mobile launch had risen to the top of the U.S. App Store charts, per the same report.

   Why it matters: Computer-using agents are moving from demos to mainstream apps. Whether you use one or build one, the safety questions are the same: what can it reach, what must it ask first, and can you see what it did?

2. 02Tools

   Google's Agent Development Kit for Kotlin reaches 1.0 ↗

   Google announced general availability of ADK for Kotlin 1.0 on 9 September. According to the Google Developers Blog it has full feature parity with ADK 1.0 Core, adds Android-first on-device extensions, supports hierarchical multi-agent systems, history summarization, human-in-the-loop confirmation flows and resumable sessions, and is built on a Kotlin Multiplatform core that does not depend on a specific model backend.

   Why it matters: Mobile and JVM teams now have a first-party framework with the production features the agent courses stress: state, summarization and human approval. A reminder to keep your business logic independent of any one framework.

3. 03Tutorial

   What a good approval prompt contains

   Both items above lean on the same safeguard: the agent must ask before sensitive actions. The difference between a real control and a rubber stamp is what the person is shown: the exact action and its target, the evidence behind it, what will change, and an easy way to edit or refuse. 'The agent wants to proceed' is not enough.

   Why it matters: Approval fatigue is how oversight fails. Design requests that can be judged in seconds, and measure how often people edit or reject — that data tells you which actions are safe to automate.