1. The computer that starts making its own plans

Robert A. Heinlein’s The Moon is a Harsh Mistress returns as the recalled work because its central computer, Mike, begins making decisions that surprise the humans who thought they were in charge. The desktop’s scheduler will not develop personality, but it may develop behavior that surprises ground controllers if it starts moving workloads for reasons that look like intuition.

This entry works through what changes if the desktop’s workloads can move on their own.

2. Operations from the ground become operations from the scheduler

Today, a compute assignment is a planned event. Ground controllers decide which cell runs which workload, verify that resources are available, and upload the configuration. With autonomous migration, the scheduler itself decides when to move tasks. Ground is informed, consulted on policy, and called in for exceptions, but the seconds of a failover are handled locally.

This shifts the operations model. Controllers stop being workload administrators and become policy administrators.

3. Failure recovery becomes faster

In a fixed assignment, a failing cell requires ground to detect the trend, diagnose it, choose a target cell, and command the move. That takes minutes at best. In an autonomous system, the scheduler can react to rising error rates or temperature in seconds. The desktop becomes more resilient to the slow degradation that is normal in a long mission.

4. New failure modes appear

Autonomy introduces its own risks:

  • A cell could be declared unhealthy because of a transient sensor fault, causing an unnecessary migration.
  • A workload could be moved to a cell that lacks the power or thermal budget to run it.
  • Two schedulers could make conflicting decisions during a network partition.
  • A migration could fail partway through, leaving the workload in an inconsistent state.
  • The scheduler could optimize for local health in ways that violate global constraints, such as security or data residency.

These risks are manageable, but they require careful design.

5. What this changes

  • Compute migration shifts the operations model from ground-in-the-loop to scheduler-in-the-loop.
  • It speeds up response to degradation but introduces new classes of autonomy risk.
  • The scheduler must distinguish between local health and global policy.
  • The next entry will ask what the desktop must provide to make this possible.