Is This a Job for an Agent? Five Questions to Ask First

Most advice about SI agents starts after you've already decided to use one. How to set it up. What to let it touch. When it should stop and ask.

This post is about the step before that. Should this task go to an agent at all?

It's an easy step to skip. A new tool is exciting, and almost anything can be phrased as a task. But some jobs get faster when you hand them over, and some get slower, because you end up checking, fixing, and redoing what comes back.

Here are five questions I'd ask first. None of them needs a spreadsheet. Each one has a plain answer, and each answer points somewhere.

Question 1: Can you say what "done" looks like?

If yes: Good sign. "A list of every page on the site that links to the old address" is a finished thing you can recognize when you see it.

If no: Stop here for now. "Make the website better" isn't a task. It's a direction. An agent given a direction will pick its own finish line, and it may not be yours.

The fix usually isn't to skip the agent. It's to do the deciding part yourself first. Turn the direction into one task with a clear end, then ask again.

Question 2: Can you check the result faster than you could do the work?

This is the one I'd weigh most.

If yes: The agent saves you time even if you look over everything it hands back. Labeling a folder of old files by topic is slow to do and quick to check.

If no: Think twice. If checking a summary of a long report means reading the whole report anyway, the agent hasn't saved you the reading. It's added a step.

Sometimes that's still worth it, if a second pass catches things you'd miss. But be honest about which kind of task it is. Work you can't check is work you're taking on trust.

Question 3: If it gets something wrong, can you undo it?

If yes: Lower stakes. A draft can be rewritten. A copy of a file can be thrown away.

If no: Hand over the preparing, and keep the doing. Let the agent draft the email, and you send it. Let it list the files to delete, and you delete them. I wrote more about where those lines go in Tell Your Agent When to Stop and Ask: https://johndclay.com/blog/agent-stop-and-ask

This question doesn't usually say "never." It says "split it."

Question 4: Will you do this more than once?

If yes: Worth the setup. The first time you hand a task to an agent, you spend time explaining it, watching it, and fixing the instructions. The second time, most of that is already done. By the fifth, it's routine.

If no: Compare honestly. For a small one-off job, explaining it and checking the result can take longer than just doing it.

There's a middle case: a one-off job that's large and dull. That can still be worth handing over. Just know that you're paying the setup cost once and not getting it back.

Question 5: Does it depend on something only you know?

If yes: Write it down first, or keep the task. An agent knows what's in its instructions and what it can read. It doesn't know that a certain reader prefers short emails, or that a file was left alone on purpose.

If no: Fine. Plenty of tasks are mostly mechanical, and the judgment they need is the kind you can spell out.

If the answer is yes and you can write the missing piece down in a few lines, do that, and the task often becomes a good one to hand over. If you can't put it into words at all, that's a sign the task is really yours.

Reading the answers together

You'll rarely get five clean answers. Here's how I'd read a mixed set.

• Yes to 1 through 4, no to 5: Hand it over. Check the first few results closely, then less often.
• No to Question 1: Not ready yet. Decide what done means first.
• No to Question 2: Keep it, or hand over only a piece you can check.
• No to Question 3: Split it. The agent prepares, and you do the final step.
• No to Question 4: Only if it's big. Otherwise, just do it yourself.
• Yes to Question 5: Write down what only you know, then ask again.

Often the best answer is "split it." The agent does the long, checkable part. You keep the part that's hard to undo or hard to explain.

A quick example

Say you want to clean up an old folder of drafts.

1. Done looks like: every draft is marked keep, archive, or delete. Clear enough.
2. Checking a list of file names and labels is much faster than opening every draft yourself. Yes.
3. Deleting can't always be undone. So the agent labels, and you delete.
4. You'll probably do this again next season. Worth the setup.
5. Some drafts are kept for reasons only you know. So you write one line: "Anything with 'keep' in the name stays."

Result: hand over the sorting and labeling, and keep the deleting. That's a good job for an agent, with one step held back.

Where this fits

This is the first question, not the last one. Once a task passes, the next questions are about what the agent can touch and when it should stop.

I'd start it on read access only: https://johndclay.com/blog/agent-read-access-first

Then give it plain stop rules, and have it leave a handoff note at the end of each session: https://johndclay.com/blog/agent-handoff-note

The short version

Before you hand a task to an agent, ask five things. Can I say what done looks like? Can I check it faster than I can do it? Can I undo a mistake? Will I do it again? Does it depend on something only I know?

Five short answers will tell you whether to hand it over, split it, or keep it.

When I'm not writing about SI, I'm writing fiction. The Shadows Series is hard dystopian science fiction about memory, identity, and the people who refuse to disappear. The series page shows where it starts: https://johndclay.com/books