Grex runs bounded tasks, called missions, with AI agents on computers the operator owns. A small control plane stores those missions and sends work to the machines, and it runs no AI itself.

On October 28, 2026, a Linux machine filled its disk while running a single mission. That day, Grex added health monitoring, cleanup and safe slots so a machine can run several missions at once.

One mission per machine left no room for anything else

Until October 28, 2026, every machine ran a single mission at a time. A review also occupied a whole machine, and nothing else could run there. While one mission was building and another machine was reviewing it, nothing else could run.

What the full disk exposed

Finished missions left their files on the machine, and nothing removed them. Grex could not see how much of the machine’s disk was in use. The incident exposed two separate gaps: finished files were never cleaned up, and disk usage was invisible.

Health readings for every machine

Each machine now reports a set of health readings back to Grex. The dashboard shows every one of them as ok, warn or critical. Seven days of history are kept for every reading on the dashboard.

Disk, memory and processor readings
Linux memory pressure and disk pressure
Network round-trip time

Cleanup for finished work

Grex now removes the files that finished missions leave on the machine, and download caches also have size limits. Cleanup also removes records for older missions once they are seven days old.

Slots that only open on healthy machines

Each machine can now run several missions at once instead of one. The operator decides how many parallel slots each of the machines should offer. An extra slot takes new work only while that machine’s disk, memory and pressure readings are ok. When those readings are not ok, no extra work is taken on.

A memory boundary for each mission on Linux

On Linux, each mission now runs inside a memory boundary of its own. Running out of memory stops that one mission instead of its neighbors on the same machine. This memory boundary was tested on a real Linux system before release.

A third machine and the first parallel run

A third machine joined the same evening, and the three machines now offer ten slots in total. The first live test ran two missions at once on one machine.

Two slots on the Linux machine
Four slots on each of the two Windows machines

What is still unmeasured

Grex has not measured whether parallel work finishes more total work than one AI coding tool working alone. The slot counts themselves are not yet tuned from any measured data. This post reports the change and the first parallel run, and it claims no measured gain.

Grex machines now run several missions at once