Concept
What is an AFK agent?
An AFK agent is an AI coding agent that runs tasks, verifies its own work, and reaches a clear terminal state — without requiring a human at the keyboard. When it finishes, it notifies you asynchronously. When it’s stuck, it stops and asks instead of guessing.
“AFK” comes from gaming — Away From Keyboard. In the context of AI coding agents, it describes a specific design posture: the agent is built around the assumption that no one is watching. That changes what the agent has to do. It needs terminal states so it knows when to stop. Self-verification so “finished” means “checked.” Reversibility gates so unattended mistakes are bounded. And a communication channel that reaches you after the fact, not during.
We built Agent AFK around this idea and coined the term. We think it names something real that existing vocabulary misses — the gap between “coding agent” and “fully autonomous.” This page explains the concept as we understand it.
Where AFK agents sit on the autonomy spectrum
Not every AI coding tool works the same way. The difference is how much of your attention they need.
- ChatbotYou ask, it answers, it waits.
- CopilotSuggests code as you type. Needs you present.
- Coding agentTakes a task and acts on it. Can edit files, run commands, iterate.
- AFK agentRuns the full task, verifies the result, and reaches a terminal state without anyone watching. Notifies you when done.
- Software factoryFully autonomous. Takes tickets, ships code, deploys. The human sets direction but is not in the execution loop.
AFK agent vs. interactive coding assistant
Most coding agents can run without you present. The question is whether they were designed for it.
| Property | Interactive agent | AFK agent |
|---|---|---|
| Requires you per turn | Yes | No |
| Runs overnight / headless | Fragile | By design |
| Verifies its own work | Sometimes | Structurally |
| Knows when to stop | No formal terminal state — user must verify | Terminal states: Done, Blocked, Asking, Interrupted |
| Notifies you out-of-band | No | Yes — Telegram by default, extensible |
| Explains what it did | Scrollback | On-disk trace with every step |
How an AFK agent works
Running without supervision is an architectural decision, not a checkbox. These are the properties that have to hold for unattended execution to be safe.
Terminal states
Every run ends in one honest place: Done, Blocked, Asking, or Interrupted. The agent can’t drift into an ambiguous middle ground. If it finishes, it says so with evidence. If it’s stuck, it says what’s blocking it. If it needs you, it asks one specific question and waits.
Self-verification
“Finished” has to mean “checked.” In Agent AFK, verification is dispatched as a separate subagent with no shared context — it re-derives claims from scratch, runs tests, and reviews the work independently. The agent that built it is not the agent that reviews it.
Reversibility discipline
Unattended execution is only safe if irreversible actions are gated. Small blast-radius work (edits, scratch files) runs freely. Large blast-radius work (pushes, deletes, external calls) requires verification or explicit intent. Blast radius determines gate level — the architecture encodes that constraint, not the agent’s judgment in the moment.
Asynchronous communication
The agent doesn’t assume you’re reading a terminal. It sends results through a channel you’ll actually check — a Telegram message, a structured receipt, a trace file. The handoff is designed for the gap between “you walked away” and “you came back.”
Subagent orchestration
Long tasks split into parallel subagents that can each verify their own lane and converge on a coherent result. One interrupt tears down the whole tree. Nothing keeps running in the dark.
When an AFK agent is the wrong tool
AFK agents add overhead — terminal-state logic, verification steps, async notification — that is not always worth it.
- Tight feedback loops. If you’re prototyping and want to steer every few seconds, a copilot is faster.
- Throwaway scripts. A 10-minute task doesn’t need verification workflows and Telegram pings.
- Tasks that require continuous human judgment. If every step needs your taste, context, or approval, the async handoff creates friction instead of saving time.
- Environments you can’t trust to recover. If a failed run would leave wreckage you can’t clean up, the reversibility gates may not cover your specific risk surface.
Frequently asked questions
- What does AFK mean in "AFK agent"?
- AFK stands for Away From Keyboard. In gaming, it means the player stepped away. For coding agents, it describes an agent designed to keep doing useful work while the developer is away — sleeping, in a meeting, or working on something else.
- How is an AFK agent different from a copilot?
- A copilot suggests code while you type and needs you present to accept, reject, and guide each step. An AFK agent takes a task and runs it end to end without you there — verifying the result, reaching a terminal state, and escalating to you only when it hits something genuinely ambiguous. The feedback loop is async, not per-keystroke.
- How is an AFK agent different from a fully autonomous software factory?
- A software factory removes the human from the loop almost entirely. An AFK agent still escalates to you when it hits something genuinely ambiguous or irreversible — it reaches a Blocked or Asking state and waits. The human stays in the loop for judgment; the agent handles the execution.
- Is Agent AFK free?
- Yes. Agent AFK is open source under Apache 2.0. The full runtime is free with no feature gates. You bring your own model provider (API key). Paid plans cover ongoing support from the maintainer, not access to the software.
- What model providers does Agent AFK work with?
- Any provider that speaks the Anthropic or OpenAI API format, including local models served through compatible endpoints. Your keys, your machine, your rules.
- Can I trust an AFK agent to run without me?
- That’s the design constraint the whole architecture is built around. Each safeguard covers a different failure mode: reversibility gates prevent irreversible damage, terminal states prevent silent drift, verification workflows prevent fake completions, and trace history lets you see exactly what happened. If the agent is unsure, it stops and asks instead of guessing.
Try Agent AFK
Everything described on this page ships in the open-source runtime. Free, Apache-2.0, bring your own model.