Files
derez.ai/blog/ideas/001-why-backing-up-an-agent-is-so-hard.md
T

3.3 KiB
Raw Blame History

001 — Why backing up an agent is so hard

Status: Research Tags: Technical, Backups, DevOps

Research notes

Real filesystem changes — from clean start through all speed-runs

Snapshot 1: After one week of basic work (original data)

504.6M  ./usr
 43.3M  ./var
448.9M  ./root
 54.0K  ./etc
996.8M  total

Home-dir-only backup would capture 448.9M (45%) — already missing 55% of what changed.

Snapshot 2: After all speed-run setups including competitor profiling (current)

  4.3G  ./usr       ← 8.5× growth from 504.6M
 487.0M  ./var       ← 11.2× growth from 43.3M
  2.1G  ./root       ← 4.7× growth from 448.9M
  1.9M  ./etc        ← 35× growth from 54K
  6.8G  total        ← 6.8× growth from 996.8M

Home-dir-only backup now captures 2.1G (31%) — missing 69% of what changed.

What accounted for the growth (drill-down)

Location Size What Speed-Run Trigger
/usr/local/lib 2.3G Hermes runtime, node_modules, python3.11 libs Hermes Agent installation
/root/.npm/_cacache 390M npm package cache Node.js tooling for Hermes
/root/.hermes 412M 19 skills, 1 plugin, 2 cron jobs, memories, config All speed-runs
/root/.cache/huggingface 142M HuggingFace model weights STT / model download
/root/.cache/uv 186M UV Python package cache Hermes venv management
/var/cache/apt 379M Debian package cache System dependencies
/var/lib/apt 80M APT package state System dependencies
/usr/local/bin 28M Executables (hermes CLI, etc.) Hermes Agent installation
/root/.config 475K App configs Misc tools
/etc 1.9M System config (hostname, apt sources, etc.) OS configuration

Total blog content created: 5 posts (Odoo, Hermes setup, competitor profiling, cold email, mobile.de research)

What this means for the post

  • Agent environments are not just config files — they're full Linux systems with packages, services, and state scattered everywhere
  • "Backup your home directory" is dangerously incomplete advice for AI agents
  • The sprawl gets worse over time: after one week, 55% outside /root; after 3 speed-runs, 69% outside /root
  • Each new skill, plugin, model download, or cron job adds state in a different directory
  • Derez.ai's full-disk snapshot approach is the right solution — it captures everything, not just /root
  • This is a strong selling point: the agent works after restore, not just the config

Outline

  1. The sprawl problem — why agents are harder to back up than a standard server
    Include the dual snapshot comparison table (Week 1 vs After Speed-Runs)
  2. Real data from a real agent — walk through the 6.8G of state and where it lives
  3. The 69% problem — what you lose with a home-dir-only backup
  4. How Borg/deduplicated snapshots solve it
  5. How Derez.ai does it automatically (one-click restore, full-disk snapshots)
  6. Better save than sorry — the selling point with real numbers

References

  • Borg backup docs
  • Hermes Agent directory structure (~/.hermes/)
  • Speed Run: Competitor Profiling (derez.ai/blog/posts/speed-run-competitor-profiling.html) — the speed-run that pushed the system past 1G
  • Full filesystem analysis: du -sch /usr /var /root /etc before and after speed-runs