Aulafy
ESEN
Courses
Courses/Agents and automation/24/7 offline agents on real hardware
Lesson

24/7 offline agents on real hardware

A local 24/7 agent doesn't fail only because of tokens: it fails because of heat, RAM, VRAM, reboots, slow model loading, full disks, and tasks nobody stops. Design for operations before you sell yourself on autonomy.

  • Choose architecture based on Mac, GPU PC, mini PC, or VPS.
  • Separate inbox, worker, and model.
  • Create watchdog, logs, and resource limits.

Honest architectures

  • Mac M-series: comfortable, quiet, and a good dev environment; watch prefill and unified memory.
  • RTX 3090/4090: more VRAM and CUDA; watch heat, power draw, and drivers.
  • Mini PC: useful for orchestration, not for large models.
  • Cheap VPS: fine for n8n/queues; bad for heavy inference without a GPU.
TerminalCode
agent-runtime:
  inbox: sqlite_or_postgres
  worker_concurrency: 1
  max_runtime_minutes: 20
  max_steps: 12
  idle_sleep_seconds: 10
  healthcheck:
    - ollama_api_alive
    - disk_free_gt_10gb
    - memory_free_gt_2gb
    - no_repeated_error_3x
  on_failure:
    - pause_task
    - save_trace
    - notify_human

Local operations checklist

  • One worker to start.
  • Persistent queue so tasks aren't lost on reboot.
  • Per-task logs, not just global logs.
  • Final summary with commands, files, and errors.
  • GPU/CPU/time limit before retrying.

Official sources

  • Ollama API
  • LangGraph Persistence
  • n8n queue mode
  • Docker Compose
Project: 24/7 agent with inboxOOM and memory management