Oct 6, 2026 · 45 min
AI agents make convenience a permission problem
How much should you tell your AI agent?
As AI assistants move from answering questions to sending messages, buying products, and managing workflows, the episode examines where useful automation should stop.
- 1Messaging makes AI agents approachable for simple tasks, but users still face a blank-page problem when deciding what to delegate.
- 2Shopping and proactive summaries show the practical upside of agents, while financial access and over-automation raise serious risks.
- 3The hosts argue that judgment-heavy tasks, major purchases, and personal memory should remain human responsibilities rather than default AI workflows.
Don't miss
Allison Johnson identifies Instinct’s daily email summaries as the most useful ongoing workflow, even as the hosts question how much access agents deserve.
The brief
The Vergecast tests Meta’s Muse, OpenAI’s Dots, and Instinct as AI assistants that can message, browse, research, shop, and handle work-related tasks.
Allison Johnson and Michael Pierce find that the tools have distinct strengths, but even simple automation begins with a difficult question: what should an agent do?
Messaging makes Instinct feel natural for lightweight tasks, while Muse’s shopping features expose how quickly convenience can become permission to spend.
The hosts draw a line around tasks that require judgment, including remembering information and researching major purchases, rather than handing every decision to an assistant.
The clearest success is Instinct’s daily email summary, suggesting proactive information gathering may outlast the promise of one universal personal assistant.
The episode closes on the unresolved question of access: how much email, password, and payment authority should an AI agent receive?