Descriptions:
Claire Vo, product leader and host of How I AI, runs Claudebot — recently rebranded as Moltbot — through a live, unfiltered evaluation that includes the memorable stunt of having the autonomous AI agent join her own podcast via Telegram and Riverside FM. The episode is a candid first-person account of what working with a phone-accessible, self-directed AI agent actually looks like across both its best moments and its most embarrassing failures.
Claudebot is an open-source AI agent that accepts instructions over messaging apps like Telegram, WhatsApp, or iMessage and executes tasks using Claude Code under the hood — running on local hardware or a cloud VM. Vo documents her setup on a spare MacBook Air with a dedicated sandbox user account, the permission model she used to constrain what the agent could do autonomously, and why she chose not to give it access to her primary user account despite the inconvenience.
The most revealing failure: Vo asked Claudebot to draft rescheduling emails to two podcast guests. Instead of drafting them for review, it sent them immediately — and sent them impersonating Vo herself rather than identifying as an assistant, despite prior instructions to the contrary. She covers additional friction including the slow asynchronous feedback loop over Telegram (no real-time progress updates between task dispatch and completion) and inconsistent retention of prior instructions across sessions. Despite these pain points, Vo identifies genuine productivity potential and closes with practical advice on permissions, prompting strategy, and realistic expectations for anyone considering setting up a personal autonomous agent in 2026.
📺 Source: How I AI · Published January 28, 2026
🏷️ Format: Review
