Ten Agents Waiting for Input: AI Coding and Hermes

Mikov says slow models have pushed him to work on several tasks in parallel, sometimes five at once. The host describes managing ten agents from a phone and choosing which one to answer based on how easy its question seems. Both find the resulting context switching painful.

Mikov says agents understand JavaScript engines, JITs, assembly languages, and optimization techniques. They do better on focused tasks with some structure, while asking a model to design something huge can send it off course. He values their persistence on boring tasks.

Mikov describes both enjoying quick AI-assisted projects and feeling uneasy when they require little expertise. He thinks job loss is a more immediate concern than AI deliberately harming people. Asked whether a nonexpert could prompt an agent to build a JavaScript engine for inexpensive watches, he says that may already be possible.

Mikov says a new Hermes release is coming soon and introduces Hermes X, a channel for prereleases that avoids waiting a year for official releases. He says he typically uses one agent per task, writes careful specifications, iterates between Codex and Claude, and checks tests closely.

Mikov says he is still adding value to agents, though he was more needed the previous year as the tools have improved. He sees humans deciding what is valuable and useful as a continuing role, but wonders how many people will be needed. The conversation closes by returning to the burden of orchestrating agents and switching contexts.