Resource pressure and host safety
Agent workloads can consume CPU, memory, processes and disk together. Keeping SSH and the control services responsive is part of operating a development host.
Reserve headroom
Do not allocate all host CPU or RAM to workloads. Reserve space for the kernel, SSH, management, updates and emergency recovery. Several individually valid jobs can still exhaust a machine unless their aggregate consumption is bounded.
Managed-host policy uses validated explicit budgets, provider VM limits and internal systemd/cgroup enforcement. Control services and workload aggregates have separate limits. A BYO administrator must establish an appropriate policy for their own host; the prerequisite script does not silently reconfigure all existing workloads.
Observe before admitting more work
Fresh observations distinguish available capacity, degraded pressure and saturation. When limits are missing, drifted or unreadable, safety-sensitive admission pauses instead of assuming the host is healthy.
Saturation should produce a typed reason and queue or reject new work. It is not evidence that running agents finished. Disconnection alone never changes an unknown process into a successful task.
Recover deliberately
Keep a recovery SSH route and enough process and memory headroom to use it. Inspect the specific workload and its resource scope. Stop only an identified abandoned workload; do not kill the daemon or broad groups of unrelated user processes as a generic repair.
Check disk and caches as well as RAM. Use bounded reusable caches and retention. Temporary scratch should be cleaned after a run; deleting arbitrary project directories is not cache management.
Updates and restart storms
Restart limits, stop timeouts and OOM handling should prevent a failed service from spinning indefinitely. Signed updates need sufficient headroom to finish verification and activation. A tight limit that prevents recovery is not a safe default.
After pressure subsides and fresh observations confirm the expected policy, admission can recover. An old cumulative OOM counter should not pause a host forever.
Pilot boundary
Hosted task execution is disabled in the owner test, and complete workload cgroup ancestry remains a launch gate. Do not interpret the presence of resource-policy contracts as proof that every vendor-launched process is already confined on your own host.