Hand your agent a half-day of work and walk away

Most CEOs hand an AI a five-minute task and then hover over it. The operators getting real leverage hand it half a day of work, set one checkpoint, and go to a meeting. That handoff is a skill, and you can learn it this week.

Peter Steinberger, who built the OpenCLAW harness, said something on X that stuck with me: "Yielding agents is a skill." He reports that once he set his agents up to run with less hovering, his tasks went from 30-to-60-minute jobs to ones that run 4 to 10 hours, and his confidence in the result went up, not down.

That matches what I've seen running Claude Code and OpenCLAW across Headphones.com and Lantern.is. The leverage was never in waiting for a smarter model. It showed up the week I stopped babysitting small tasks and started handing over big ones. Now I'll point an agent at a full competitor sweep, or a first pass at the quarterly board narrative, kick it off before a meeting, and review what's waiting when I walk out. The work happens while I'm somewhere else. That is the whole trade.

The short answer

Scope the task big, then gate the result. Pick a job with a finish line you can check, write the AI agent a brief so it knows your business, put your review at the end instead of over its shoulder, and let it run while you do something else. The skill is not in the prompt. It is in choosing a task that can survive an hour alone, and setting up the one checkpoint that catches a bad result before it ships.

The full walk-through

Step 1Pick a task with a checkable finish line

The first jobs you hand over should have a clear "done." A competitor scan that ends in a one-page summary. A first draft of the board-deck narrative. A cleanup pass on every open deal in the CRM. You want a task where you can look at the result in five minutes and know whether it worked.

Avoid open-ended mandates on day one. "Improve our marketing" has no finish line, so the agent drifts and you can't tell if it succeeded. "Rewrite these six landing-page headlines and tell me which one performed best last quarter" has a finish line you can check at a glance.

Step 2Write the brief before you hand over the keys

An agent that has never been told who you are will guess, and guesses compound over a long run. Give it a short brief about your business first: who your customers are, what your terms mean, what good looks like, what it must never touch. In Claude Code that brief lives in a CLAUDE.md file. Give it persistent memory too, so the next long task starts from what the last one already learned.

This is the difference between an agent that runs for an hour and stays on the rails and one that quietly wanders off them. (Here is why long tasks drift in the first place.)

Step 3Put the review gate at the end, not over its shoulder

Hovering is the thing that keeps your tasks small. The fix is a gate: a single point where you check the work before anything irreversible happens. Steinberger runs an automated review pass that reads the agent's work before it lands and flags problems; he says it catches edge cases he would have missed and sometimes runs for hours on its own. You don't need his exact setup to get the idea. Put yourself, or a second agent, between the work and anything that sends, pays, publishes, or deletes.

Get Claude to summarize what it changed and why at the end of the run. You review the summary, not every keystroke.

Step 4Start it, then leave

This is the part that feels wrong the first few times. Kick off the task and go to your next meeting. The whole point of a half-day task is that the agent works while you don't. If you sit and watch, you've turned a four-hour agent task back into a four-hour you task, which is the opposite of the trade you're making.

Step 5Review like an editor, not a co-author

When you come back, read the result the way an editor reads a draft. Is the finish line met? Is anything wrong that actually matters? Send it back with specific notes, or ship it. You are judging the output, not re-doing the work. The first time half a day of work is sitting there waiting for your sign-off, the skill clicks.

When it breaks

The task was too vague. If the agent wandered, the brief or the finish line was fuzzy. Tighten the "done" and re-run. A bounded job that fails is fixable; a sprawling one just drifts.

There was nothing to check. If you can't tell in five minutes whether it worked, the task had no checkable output. Add one: a file, a ranked list, a recommendation with a reason. The output is what your gate inspects.

You hovered anyway. If you watched the whole run, you didn't hand anything over. Pick a lower-stakes task next time, one you trust enough to walk away from, and build up from there.

Want to maximize your AI leverage? Upgrade to Pro.

Pick one task on your plate this week that has a clear finish line. Write the agent a one-page brief, set your checkpoint at the end, start it, and go to your next meeting. Tell me what was waiting for you when you got back. I love hearing which job people hand over first.

Related

• Why most AI agents fall apart in real work
• What is an AI agent?
• What is a coding agent?
• What is a harness?
• Persistent memory across Claude Code sessions
• What is a CLAUDE.md file?