AI CompendiumA public chart of the language
Patterns of Work

AFK

also called Away from keyboard, Unattended

Running without a person watching.

Definition

Agentic work executed with no human present — overnight runs, queued tickets, background jobs. The opposite of interactive work.

AFK changes the design constraints entirely: every ambiguity must be resolved up front, failures must be recoverable without a human, and the blast radius must be contained in advance.

Unattended work succeeds or fails on what you set up before leaving. Three things have to be true: the task is bounded, the acceptance criteria are machine-checkable, and failure is safe and visible. Where any of those is missing, a person sitting there is not a bottleneck but the control.

How people use it
Deciding what can run overnight.
"Only AFK-able work goes in the queue: mechanical, well-specified, inside the sandbox, with tests as the check."
Worked examples

The overnight job that reported back

A batch of small improvements is queued for an unattended run.

Each is scoped to one file, must leave the build green, and opens a branch rather than merging. In the morning there is a set of reviewed candidates and a list of the ones that failed, which is a better outcome than either total success or a broken main branch.

The work that should not have been delegated

An authorisation refactor is queued as an unattended task.

Security-sensitive, ambiguous, and the failure mode is silent. This is the category the guidance explicitly excludes — the cost of review afterwards approaches the cost of doing it, with the risk of the work being plausibly wrong.

Avoid

Queueing ambiguous work for unattended execution and being surprised by what comes back.

Sources
  • blogEffective harnesses for long-running agents ↗Anthropic Engineering · 2025

    The initializer-plus-incremental-sessions pattern for work that outlives a context window, and the four ways it fails: doing too much, losing context between sessions, declaring victory early, not testing end to end.

  • docsClaude Code overview ↗Claude Code docs · 2026

    A first-party description of the surfaces an agentic coding tool ships on, including parallel sessions and long-running tasks. Not evidence that it improves productivity.

  • docsBest practices for using GitHub Copilot to work on tasks ↗GitHub docs · 2026

    Recommends bounded tasks with explicit acceptance criteria, and names what is a bad fit for unattended delegation: ambiguous, security-sensitive, production-critical or deeply domain-specific work.

Related terms