Why WorkBuddy feels dumb: ten settings that keep agents alive, safe, and worth trusting

Desktop agents that can touch local files fail or become dangerous because of their defaults: a locked screen can kill a job, storage can overflow, and every action prompt can train users to click through. These settings define what unattended agent work safely requires, and the same checklist applies to any local-agent tool a team evaluates.

WorkBuddy is a desktop agent that can read files and execute commands, so its default behavior has sharp edges. Long tasks stop when the screen locks if 'continue after screen lock' is off. Data, logs, session history, downloaded materials, and skills accumulate under C:\Users\<user>\WorkBuddy, which can squeeze a small system drive. The experience improvement program sends anonymized usage data to the vendor unless disabled. Reply style presets change tone without changing model capability, and custom instructions let standing preferences be read before every task.

Memory features import a profile from another AI like GPT through an official prompt, support editing stale memories, and carry context across conversations. Skill auto-updates and trusted-skill auto-install keep accumulated abilities current without manual checks. Linking the desktop client to a WeChat mini program and mobile app enables remote monitoring and mid-task instructions. External models can be plugged in through API keys, coding or token plans, or free models. The security center is where the control sits: sandbox rules govern file, command, and network access; automatic backup keeps originals before modifying documents, spreadsheets, code, or batch files. One practical stance is full access for routine work to avoid approval fatigue, with stricter permissions restored for system-level changes.