← All postsAI Cloud Agents 
Agents need more than a text model
A model that can only emit tokens cannot verify its own work. Giving agents a real machine — with a filesystem, a network, and a screen — changes what they can finish.
Sandal Engineering6 min read
An agent that writes a patch but never runs the test suite is guessing. The gap between plausible output and correct output closes only when the agent can execute, observe, and retry.
Execution as a primitive
That means a real filesystem it can write to, a network it can reach out from, and a process tree it can inspect — not a sandboxed function call with a thirty-second ceiling.
Beyond Linux
Plenty of work does not happen on Linux. Signing an iOS build needs macOS on Apple silicon. Testing a desktop installer needs Windows. An agent limited to one operating system is limited to one slice of the job.