Agent infrastructure gets real: sandboxes, runtimes, and covert channels

Nvidia's OpenShell, a runtime for fleets of autonomous AI agents, was all over the demos at GTC Europe, and commenters were quick to flag its default telemetry and how to turn it off. Jank, a native Clojure dialect with C++ interop, drew the comment that REPL-driven development with access to native code could be 'a magic formula for agent engineering' with no restarts or recompiles. Meanwhile a thread on how Fable 5.1 hides messages in images raised the uncomfortable follow-up: agents could hide content in ways operators can't spot from output or reasoning traces.

The pattern: the interesting work is moving from the model to the box around it. Who isolates the agent, who sees what it does, and what it can smuggle out are the open problems. The small WorldTUI experiment, a 360M model trained to act like a terminal, shows how cheap fine-tuning on narrow tasks has become.

The skeptics ask what the point is when you already have a terminal. Novelty projects still outnumber production uses.