Machine Guard

Running multiple agent-driven and long-lived processes on one machine risks contention no single project can see on its own — a machine can be starved of resources by the combination of workloads even when each one looks reasonable in isolation.

observe (aggregate resource pressure)
   → attribute (which known workload is contributing)
   → record bounded local evidence
   → advise the project that owns the workload

Machine Guard runs continuously as a background service. It does not act on what it observes beyond surfacing it.

• Machine-level, cross-project view — resource pressure is judged in aggregate, not per project
• Observation, attribution, and evidence are kept strictly separate from control: it advises, it does not kill, throttle, reboot, or otherwise mutate a workload
• Goal-preserving escalation — the owning project decides how to respond, rather than having that decision made for it
• Bounded, local evidence retention rather than open-ended logging

Active — running as ongoing infrastructure. The backing repository is private, so there is no public repository link at this time.