Case study: a household of agents

Описание и применение

Автор создал домашнюю систему AI‑агентов, которая обеспечивает надёжное выполнение задач через детерминированные скрипты, плагины‑ограничения и внешние проверки, что повысило надёжность и контроль над автоматизацией.

Задача

The author needed reliable follow‑through for a constant stream of small obligations but struggled with agents that talked about work instead of completing it and failed to enforce boundaries.

Результат

Over six months, 77 scheduled jobs were defined, 14,563 runs recorded with 13.6% incomplete, 6,308 model turns (~900/day), and 24 enforcement plugins enabled. The system reliably delivered reminders, verified archives, and automated tasks, while eliminating unsafe agent behavior.

Рабочий процесс

• 1. Install platform with cheap, always‑on models. 2. Build agents with dedicated shells. 3. Add enforcement plugins to block unsafe actions. 4. Implement memory with schema, size cap, and per‑topic files. 5. Replace prompt‑based routing with lookup tables. 6. Add verification artifacts and presubmit scripts. 7. Use hooks instead of prompts for agent actions. 8. Cap blast radius with mechanical rules. 9. Apply same discipline to builder with hooks and snapshots. 10. Monitor and iterate based on logged runs and metrics.

Доказательства

The short version is uncomfortable. The value never came from how smart the models were. It came from how fast I could turn a lesson into machinery: a script, a state file, a cron job, a hook, a checksum.

Источник

Verification that lives outside the agent. An agent's report about its own work is a second artifact that may or may not match the first.

Источник

Hooks instead of prompts. The comment at the top of the first enforcement plugin states the whole problem: the platform can restrict what agents can't do, but cannot force what they must do.

Источник