Agents need sandboxes, and the human-in-the-loop part is awkward

MXC launched as a sandboxed code execution system. The top reaction asked why everyone is building their own agent runtime when a sandboxing policy design, like Unix permissions, should come first. Defenders called it a friendlier frontend over bubblewrap, seatbelt and process containers, which are painful to configure consistently. A separate post lets agents draw big arrows, boxes and text on your screen. Its pitch is that agents "keep hitting the same wall, the part that only a human may do." The reply was blunt: if you're just there to click sudo for the bot, give it root and be done.

On the lighter end of the same wave, people used Claude Code to find a possible new planet, port Quake to safe Rust (the first question was whether an LLM did it), and sketch a Home Assistant dashboard from a floor plan. A Show HN also shipped OSC 7501, a terminal protocol for program status, so tools can report what they're doing.

The through-line is that the agent runtime layer is unsettled. Everyone is hand-rolling isolation, permissions and status reporting, and nobody has settled how a person fits into the loop without becoming a rubber stamp.