The Unattended Coding Agent: What 'Done' Actually Means

Unattended does not mean "set it and forget it." It means you can walk away from the session and come back to something reviewable.

Most AI coding setups call themselves unattended because you can start a long task. They are not unattended in any meaningful sense — you still have to watch for failures, catch hallucinated tests, and manually verify the output before you can trust it. That is not unattended. That is unsupervised, and those are different things.

The Three Requirements for a Genuinely Unattended Setup

Every unattended coding run that ends well has three things in common. If any one of them is missing, you are not running unattended — you are just hoping the agent does not trip.

1. Bounded scope. The task has a spec, not just a prompt. A prompt is loose and open-ended. A spec defines what done looks like, what files are in play, and which tests must pass. Without that boundary, the agent will drift — and because you are not watching, you will not know until you come back.

2. Automated verification. Something checks the output before you see it. If the only verification step is "I'll look at it in the morning," you are the verification step. That means you are still attending. Automated verification means tests run, diffs are checked, and the run either passes or loops back — without you.

3. Clean re-entry. If it fails, you know exactly where and can resume without starting over. The worst outcome from an overnight run is not a failure. It is a big diff with a confident summary and no clear way to tell what the agent actually did, what it skipped, and whether any of it is mergeable. Clean re-entry means the run leaves a trail you can follow.

The Overnight Test

The real test of an unattended setup: can you start it at 11pm, sleep 8 hours, and wake up to something you can actually review?

If your current setup cannot pass that test, it is not unattended — it just does not require constant input. Those are different things, and the distinction matters most when you come back to the result the next morning.

How Ralph Workflow Makes This Concrete

Ralph Workflow is a composable loop framework that encodes all three requirements into a repeatable workflow. The default workflow runs four phases — plan, develop, verify, deploy — with explicit finish-state checkpoints at each gate.

spec-first → agent builds → verify catches mistakes → loop if broken → clean output

You write the spec once. The orchestration loop handles the rest. The verify step is what catches the agent's mistakes before you ever see them. And because every phase produces a concrete artifact, re-entry is never guesswork — you know exactly what ran, what passed, and what still needs attention.

This is not a theoretical architecture. It is a practical shell: the same loop structure that works for a three-file refactor with tests also works for a cross-module migration, because the phases are the same.

Try the overnight test yourself. Ralph Workflow is free and open-source on Codeberg — star, fork, and open issues there. GitHub mirror here.

→ AI Coding Tools Compared 2026 — see which tools are actually built for unattended work and which are interactive-first

For a concrete walkthrough, start with the first overnight task guide.