Software is collapsing into one input window, but code still has to check the model’s work

Chat assistants are proliferating, with users adopting tools like GrokBot in messaging apps, Perplexity on their home screens, or wrappers in Telegram and Slack. Promises are made that typing to a bot will handle work. However, when real company operations are attempted through standard consumer assistants, a wall is hit within 10 minutes.

These bots operate in closed browser sandboxes, unable to access actual infrastructure, run headless browsers through multi-step sessions, execute Python, or inspect server jobs over SSH. The author sought a solution and moved their daily work to three components: a cheap Linux VPS, a Telegram chat, and Hermes, an open-source agent runtime.

This setup revolutionized their mornings and influenced the development of Lobby, a creator discovery platform. The author demonstrates how operations can be run through a Linux VPS, showcasing Hermes' ability to scrape websites, extract case studies, draft messages, and perform various local coding tasks. The author argues that standard SaaS dashboards feel cumbersome compared to this single input window approach, where a user types a campaign brief and the system handles the searching, checking, and verification.

However, the author warns that raw language models can fail in production, using the example of a skincare brand where an agent incorrectly identified a creator due to outdated information. To address this, the author implemented strict code validation and filters, ensuring the language model only handles semantic parsing and drafting text.

This shift represents the end of traditional dashboard interfaces, as software increasingly collapses into a single input window.