Agent AFK
Agent AFK is an open-source runtime designed for running coding agents without constant supervision. It is fully local-first, free forever under the Apache-2.0 license, and works with any model provider, including local models. Once installed via npm, users can start a coding task and walk away; the agent performs the work, verifies its own results, and sends a notification via Telegram when finished. The user returns to a detailed receipt showing every step, decision, and any pending items, so nothing is left to guesswork. The architecture is built around the six common failure modes of autonomous coding agents: linear tunnel vision, fake done, context rot, subagent drift, bad handoff, and the babysitting tax. Each failure mode is addressed by a structural design element, such as parallel hypothesis generation via /diagnose, adversarial verification via /shadow-verify, local trace history under ~/.afk/, centralized subagent tree management with a single interrupt, explicit terminal states (done, blocked, asking, interrupted), and remote supervision with reversibility gates. This makes the system robust when no developer is actively watching. Operationally, Agent AFK offers subagent orchestration, prompt and skill routing, full tool call visibility, and verification workflows. It is reversibility-aware, moving quickly on safe edits in isolated worktrees and being more cautious on git pushes, commits, deletes, or external calls. Traces, plans, and session state persist on disk, allowing users to inspect, resume, or roll back previous runs. With 35,000 downloads last month and a simple npm install command, Agent AFK aims to provide the first coding agent experience that truly does not require a babysitter.
- Category
- AI Agent
- Pricing
- Free
Problem solved
Coding agents often require constant supervision and may report completion without proper verification, leading to untrusted results and unmanageable long-running processes. Agent AFK solves this by providing a runtime that lets developers delegate tasks and step away, with automated verification, remote supervision, and comprehensive traceability. Users get notified when work is done or blocked, and can review a full receipt of actions, making autonomous coding safe and practical.
Use cases
- Run a coding agent to implement a feature while you step away
- Delegate a bug fix and get a verified receipt when it's finished
- Use adversarial verification to ensure agent work is actually correct before merging
- Orchestrate multiple subagents for complex tasks with centralized control
- Review local trace history to understand what an agent did during a long run
- Set up remote supervision via Telegram to be notified of task completion or blocking
- Use local models or any model provider with your own keys and machine
Audience
- Developers
- Startups
- Enterprise
- Hobbyists
Pros
- Open source (Apache-2.0) and free forever
- Local-first: your keys, your machine, your rules
- Works with any model provider, including local models
- Designed for autonomous operation with remote supervision via Telegram
- Includes adversarial verification to catch false 'done'
Cons
- Requires Node.js >= 20 environment
- Low number of GitHub stars suggests an early-stage project
- May require configuration for Telegram integration and model provider setup
- CLI-based, so not suitable for non-developers
- Autonomous agents can still behave unpredictably despite safeguards
Official website: https://agentafk.com/
Canonical LambdaBase page: https://www.lambdabase.com/tools/agent-afk