What is Agent AFK?
Agent AFK is an open-source runtime designed for unattended coding-agent work, enabling developers to start a task and walk away with full confidence. It addresses the common failure modes of autonomous agents: single-hypothesis fixation, premature completion claims, unreadable long-run logs, and parallel lane divergence. By enforcing permission boundaries, terminal states, trace history, verification workflows, and Telegram supervision, Agent AFK ensures that every action is governed and reversible.
Key features include parallel hypotheses with isolated worktrees, verification by default, and comprehensive orchestration. Unlike traditional agent tools that assume constant human oversight, Agent AFK assumes you are not watching. It wires orchestration, routing, traces, verification, and supervision directly into the loop. When a run goes sideways, you can read exactly what happened: every session, subagent, and tool call is stored on disk under ~/.afk/. Inline tool call details include arguments, output, and timing.
Use cases include running coding agents with Claude, Cursor, Claude Code, or custom agents without babysitting the terminal. It is ideal for developers tired of supervising every few minutes, enabling true autonomy. Technical details: requires Node ≥20 and puts AFK on your PATH. It works with your own models (BYO model) and is local-first with an Apache-2.0 license. One command, such as /diagnose, triggers a real pipeline with parallel hypotheses, isolated worktrees, and verification. Agent AFK transforms autonomous coding from a supervised chore into a reliable, hands-off process.
Use Cases
Key Features
parallel hypotheses isolated worktrees verification by default comprehensive orchestration permission boundaries terminal states trace history Telegram supervision inline tool call details unattended executionWho is it for?
developers, coding agents, Claude users, Cursor users, Claude Code users, platform teams, automation engineers
