HhiveyBLOG

How to Schedule AI Agents for Check-Ins That Actually Help

You start an agent on a long task, go make lunch, and come back to find it finished twenty minutes after you left. It has been sitting idle ever since, waiting for someone to say "next." Or it stalled on a question nobody saw. Either way, the clock ran and the work didn't.

The fix isn't checking your terminal more often. It's learning to schedule AI agents: you give them a rhythm, so they revisit the work, report back and pick up the next piece without you hovering. Here's the method I use, why it works, and how I do it in hivey.

Why you need to schedule AI agents instead of babysitting them

Most coding agents are reactive. They do a lot when you prompt them and nothing when you don't. That's fine for a ten-minute fix. On anything longer, the agent goes idle between your prompts, and you become the scheduler without meaning to.

That costs you in three ways:

A scheduled check-in turns this around. The agent gets a nudge at a known interval, looks at its own progress, and either moves on or tells you what it's stuck on. You stop polling. The work polls itself.

What a good scheduled agent check-in looks like

A timer alone doesn't help. "Keep going" every hour just produces more of whatever the agent was already doing. The check-in needs a job. I think of it as a short standup with a checklist:

Part of the check-inWhat the agent should do
StatusSay what was finished since the last check-in, in one or two lines.
Goal checkCompare progress against the original task, not the last subtask.
BlockersName anything waiting on a person, a credential or a decision.
Next stepPick the next concrete piece of work and start it.
Hand-offIf the work is done, say so clearly and stop instead of inventing more.

That last row matters. An agent that never hears "it's fine to be done" will find something to do, and usually not something you wanted.

Writing check-in instructions agents can act on

The instruction you attach to a schedule gets read many times, often when the agent has lost some of the earlier context. Write it so it stands on its own.

A few rules I follow:

  1. Restate the goal. Don't write "continue the task." Write "We're migrating the billing module to the new API. Check progress against that."
  2. Ask for a short report. One paragraph is enough. A long report every two hours turns into noise you stop reading.
  3. Define done. "When all tests pass and the PR description is written, report done and wait."
  4. Say where blockers go. "If you need a decision from me, ask it as a single question at the top of your report."
  5. Pick an interval that matches the work. Fast iteration on tests might want a short loop. Research or a large refactor is fine every few hours. Too often and the agent never gets deep; too rarely and idle time comes back.

This works with any scheduling tool, even a plain cron job that pipes a prompt into an agent CLI. The instruction is the part that does the work. The timer just delivers it.

How to schedule AI agents with hivey

I built hivey to coordinate my agents from one terminal, with Claude Code and Codex working as a team. Each swarm has its own workspace and a coordinator that directs the agents and reports back. Scheduled check-ins wake an idle coordinator at configured times, so the nudge lands on the agent that can actually reassign work.

Start by seeing what's running and what's already scheduled. Check before you add, so you don't stack duplicate timers:

hivey swarm list
hivey swarm schedule list <swarm>

Then add a check-in with an instruction written the way the previous section describes:

hivey swarm schedule add <swarm> --every 2h "Check progress against the goal: migrate billing to the new API. Report what finished, any blocker as one question, and start the next step. If tests pass and the PR is written, report done and wait."

If a schedule stops being useful, remove it by its id from the list:

hivey swarm schedule remove <swarm> <id>

You can still step in between check-ins. Send the coordinator new context at any time. Normal messages wait until the agent is idle, so you don't cut it off mid-thought:

hivey msg send <swarm>/coordinator "Priority change: skip the refund flow for now."
hivey msg log --swarm <swarm>

And to read where things stand without opening every pane:

hivey swarm info <swarm>

If you need to hold everything for a while, hivey swarm pause <swarm> holds message delivery and hivey swarm resume <swarm> releases the queued messages.

One honest requirement: hivey runs locally on macOS and Linux. Detaching with ⌥Q (or ctrl+b, then q) leaves agents running, but the machine has to stay awake and connected. Scheduled check-ins run on your computer, not on a hosted service.

Quiet hours: let scheduled agents work at night

Once your agents check in on their own, the next problem is that they check in at 2 a.m. too. You want the work to continue overnight without your phone or speakers announcing every step.

Hivey's quiet mode handles that. While it's on, voice and notification sounds stay silent, Slack posts are held and arrive as one catch-up message when it turns off, and agents keep working. You can switch it by hand:

hivey quiet on
hivey quiet off

Or put it on a daily schedule so it turns on and off by itself, even with every hivey window closed:

hivey quiet schedule 22:30-6:45

If you only want the spoken updates silenced at night, use voice quiet hours instead:

hivey voice quiet 22-8

Put together, the pattern looks like this: a coordinator wakes every couple of hours, checks progress against the goal, keeps the team moving, and saves its noise for the morning. You read one catch-up instead of fifty pings.

Try it

On an Apple Silicon Mac:

npm install -g @hivey/cli
hivey setup

Then start a team with hivey swarm new "<task>" and add your first check-in. Setup, Linux install and everyday commands are in the docs: https://hivey.dev/docs/.