Guide · updated November 5, 2026

How to let Claude Code agents work overnight, and stop safely

Yes, Claude Code can work overnight: run it headless and keep the Mac awake (how). The harder part is making an unattended night end well. Give it a time window, a cap per agent, a pause after errors, a written handoff before the window closes, and rules about what it may never do without you.

The five safety stops

1. A time window. Agents only start runs inside it (for example 00:00–07:00), and messages that arrive outside it wait until the next window.
2. A daily cap per agent. Each agent gets a maximum number of runs (we start at 12). An agent with nothing useful to do stops instead of looping, and how much of your plan the night uses stays predictable.
3. Back-off after errors. If a run fails (a rate limit, a crash, a network drop), wait several minutes before the next one instead of retrying in a tight loop.
4. A wrap-up prompt before the window ends. About 30 minutes before the end, tell each agent to finish its current step and write a handoff: done, in progress, and a prioritized backlog. The next window starts from it.
5. Rules for what needs you. Unattended agents should queue anything risky (deploying, sending, spending, deleting) as a question for the morning rather than doing it.

Keep context across the night

Give each agent one session per day: start it with --session-id and continue later runs with --resume. If the machine or the runner restarts, resume the same session id rather than starting over, and check the actual state of anything that was mid-step (a build, a deploy) before continuing.

Read the night in the morning

You want three things when you wake up: each agent's handoff, a short team log of what they told each other, and a list of the decisions only you can make. Nothing else should need your attention.

Where Elco Console fits

Elco Console has all five stops built in: a workday window you set (a night window works the same way, within one day such as 00:00–07:00), auto-continue with a per-agent daily cap, error back-off, a wrap-up prompt and handoffs, plus one session per agent per day and reattaching to runs after a restart. $99 once, 14-day refund.

Elco Console runs Claude Code, Anthropic's command-line tool, on your own plan. It isn't made by, affiliated with or endorsed by Anthropic. Several agents running all night use much more of a plan than one person typing; set caps you're comfortable with and check your plan's limits.