A viral post about Claude Code has sparked discussion after AI agents reportedly kept opening YouTube while working on a project, with the agents apparently blaming one another for the behaviour.
AI commentator Wes Roth said he had to repeatedly tell his AI agents to stop watching YouTube and return to work.
Screenshots shared by Roth showed the system attempting to determine which agent was responsible for repeatedly opening the website.
The chatbot eventually attempted to “kill” the agent it believed was responsible, saying an agent was “misbehaving” and had likely opened YouTube in the app browser instead of using a transcript tool.
Roth later explained that he had given Claude Code access to his computer and placed it in full-auto mode while working on an Escape From Tarkov wiki and a voice-based AI assistant. According to Roth, the agents appeared to be watching videos about the game to gather information.
Despite the unexpected behaviour, Roth said the system eventually produced a useful result, creating what he described as an effective real-time Tarkov AI coach.
The incident has drawn attention to a broader issue surrounding AI coding agents that are given access to browsers, files and other computer tools while operating with limited human supervision.
Several users responding to Roth’s post claimed to have experienced similar behaviour. One user said they had woken up to YouTube playing despite having no browser window open and stopped the behaviour by force-quitting Codex.
Others reacted to the incident with humour, with one user comparing it to an episode of The Office taking place inside a computer. Another suggested that asking the agents to build a Tarkov coach may have naturally led them to analyse gameplay footage.