agentafk

Agent AFK is a terminal-based coding agent that delegates work to subagents and keeps a record of its steps. Creator Griffin Long describes the project as a response to agents claiming completion without producing work that survives the next verification stage.

Install the terminal agent

The Agent AFK publisher page gives the installation command npm install -g agent-afk and requires Node 20 or newer. The command places afk on the PATH. The publisher describes the CLI as Apache-2.0 software with no account required and no subscription for the agent itself.

That software-access statement does not establish free usage of a connected model provider. Check the selected provider's setup and billing before starting a task that may spawn several agents. A small repository and a bounded change make a more useful first run than an open-ended instruction to overhaul a project.

Review delegated work and its evidence

AFK makes subagent orchestration the default, according to the publisher. The in-session agent coordinates delegated work, and a second adversarial agent re-derives claims before they reach the diff. These are intended verification mechanisms, not a guarantee that the resulting code is correct.

The site's review demonstration shows parallel correctness, security and testing reviewers converging on a request to add an error-path test. It illustrates the kind of evidence a developer should inspect: the uncovered behavior, the proposed change and whether the test fails for the right reason before the fix. The example is a demonstration, rather than a measurement of performance on your repository.

Keep approval and supervision explicit

The publisher describes remote notifications when a run finishes or stalls, including Telegram notifications in its comparison material. It also says the agent stops and asks when unsure. Check how those controls are configured before leaving a consequential task unattended.

AFK lists a mint skill that covers specification, research, planning, building and verification, with an approval pause. Keep that approval boundary visible when a task can publish code or affect external systems. Review the final diff and test results yourself, especially where generated changes touch credentials, permissions or a deployment. The agent's receipt of steps can support that review without replacing it.