Agent Autoloop

Agent Autoloop (KKKKhazix/agent-autoloop) is an MIT open-source agent automation hub: it dispatches cron jobs, GitHub issues, and alerts to local Claude Code / Codex / ACP CLI agents, exits after a round, and only pings you via Feishu/DingTalk/Slack cards when approval is needed—no process / no tokens while waiting. The author describes a ~1650-line zero-dependency core with event sourcing, instruction/data separation, and hop-count circuit breaking.

Repo created 2026-11-03 with v1.0.0 the same day; ~★979 by this scan amid bilingual chatter. Prior notes flagged it for careful evaluation; clear product naming plus ship evidence warrants insertion.

Problem

CLI agents die when the terminal closes; keeping them warm burns tokens. Autoloop turns dispatch + human-approval suspend into an exit-friendly loop.

Differentiation

Versus agent frameworks (CrewAI/Letta, …): focuses on on-call scheduling for local CLI agents with zero-cost suspend and approval cards. Versus bare cron: ships anti-injection / circuit-break narratives.

Limits

• Requires working Claude Code/Codex locally; you own secrets/permissions.
• Star spikes ≠ proven production stability.
• Complements sandboxes (e.g. Agent37): scheduling vs isolated compute.

Who it is for

Builders already on Claude Code/Codex who want issues/cron followed overnight and human gates only for risky actions.

Getting started

1. Read the GitHub README and v1.0.0 notes.
2. Install per docs (recent Node, etc.) and start with a low-risk cron template.
3. Wire Feishu/DingTalk/Slack approval cards; verify zero-token suspend before widening event sources.