Why Single-Purpose Agents Beat One Big Automation Script: A Homelab Case Study

Designing Single-Purpose Agents Instead of One Big Automation Script

Adrien Fischer’s homelab runs thirteen independently-scheduled agents for chores like container health, certificate expiry, and meal planning. Each agent writes JSON to its own outbox, pushes notifications through one ntfy topic, and is read by a single aggregation dashboard.

Why This Matters

Engineers default to monolithic automation scripts that accumulate unrelated responsibilities over time—a bug in Trakt pagination can take down Home Assistant entity checks. Fischer’s decomposition into single-purpose agents keeps each component testable in isolation with fixture files, requires no shared credentials or deployment cycles between them, and allows retiring any agent by deleting a directory without untangling shared modules.

Key Insights

• Three-convention contract (JSON outbox + one ntfy notification channel + one dashboard) keeps integration surface small; adding a sixth agent requires no changes to the other five.
• ‘Argus Filch’ watches Docker containers for restart loops and failed healthchecks; ‘Mad-Eye’s Watch’ tracks TLS certificate expiry at warning (30 days) and critical (7 days) thresholds.
• ‘Molly’s Cupboard’ reviews Home Assistant entity list weekly for unavailable entities; ‘Dobby’s Rounds’ prunes old snapshots and state files—neither reads the other’s outbox directly.
• ‘O.W.L.s’ performs daily infrastructure audit (config drift, open ports); ‘Auror Office’ correlates those findings with auth logs and Docker posture into one security digest—no LLM in the loop.

Practical Applications

• • Use case: Homelab operators can deploy independent cron-based agents for distinct tasks (e.g., network monitoring via Marauder’s Map scanning UniFi for offline devices). Pitfall: Violating the ‘one artifact format’ convention leads to implicit cross-agent dependencies that recreate monolith complexity.
• • Use case: Engineering teams designing automation pipelines can adopt an outbox pattern with timestamped archives so correlators are explicit read-only layers. Pitfall: Allowing any agent to read another agent’s output directly creates hidden coupling that breaks when either changes.

References:

• https://dev.to/po8rewq/designing-single-purpose-agents-instead-of-one-big-automation-script-ddg