looper

Looper is a manager for agentic loops that schedules AI agents like cron jobs, allows users to watch them work in real terminals, and aims to stop burning tokens on idle polling. It inverts the typical loop by using a cheap check script and optionally a small classifier model to decide whether there is work, and only then starts a fresh agent session. This design avoids idle token burn and context rot, ensuring at most one agent per task.

The system supports cron-style scheduling with wall-clock slots, timezones, and schedule end dates, as well as manual tasks. Each run starts a fresh session, preventing context accumulation; the agent signals completion with 'looper-done', files a report, and the loop resumes. Monitoring is built in, with per-task run logs, live status, and a real terminal tab that can be watched and typed into mid-run. System notifications can be configured per task, from errors-only to every run.

Looper is compatible with any harness or environment, including Claude Code, Codex, or custom agent CLIs, and can run on Windows driving agents inside WSL, or Linux/WSL directly. Task setup is done through an editor with Schedule, Check, Classifier, and Agent tabs, templates, import/export, and agents or scripts can register tasks by dropping a JSON file in the inbox. The check and classifier are optional; without them, every slot goes straight to the agent.