From Zero to a Gated Agent Loop: Six Bootstrap Stages, a Smoke Test, and a Token Ledger

Описание и применение

Пошаговый процесс создания безопасного цикла агента с ограничением токенов, включая защиту ключей, настройку сервера, тестирование конечной точки и ведение журнала расходов.

Задача

Missing stage gates in AI agent loops allow leaked keys and uncontrolled token usage, leading to costly failures.

Результат

The workflow establishes a secure, token‑budgeted AI agent loop with automated safeguards against key leakage and budget overruns.

Рабочий процесс

• Create repo and record tool versions (Stage 0).
• Copy .env.example to .env, set permissions, and grep for leaked keys (Stage 1).
• Provision a free server, create non‑root user, restrict ports (Stage 2).
• Run a single capped request to verify endpoint and usage (Stage 3).
• Implement ledger.py to record usage and abort on budget (Stage 4).

Доказательства

The model is not the risky part — the missing stage gate is.

Источник

A repo where a leaked key is a failing test, not a bad day.

Источник