Corral kills runaway agent processes

A small open-source project called Corral tackles an unglamorous problem that can cause real trouble in AI coding workflows and automated jobs. It runs a command with a time limit, then checks that every process started by that command has actually stopped before returning.

That matters because a command can leave work running in the background. A process might detach itself, keep an output stream open, ignore a normal stop signal, or continue holding a port, file lock, memory, or CPU. The next automated run can then hang or fail for reasons that are hard to trace. Corral is designed to control the full process tree, not just the direct command it launched. On Linux, it can use cgroup v2 controls when available, or fall back to tracking processes through the system process table. If it cannot prove all processes have stopped, it returns a specific failure code.

This is a narrow tool, and it is not a security sandbox. It does not restrict file access, network access, or privileges. But it’s a useful signal: as more teams let AI agents run shell commands and tests, basic operational controls matter just as much as the model itself. You need visibility into what ran, how it ended, and whether it left anything behind. This is the kind of thing we build into an AI command centre, so automated work stays observable and accountable.