Run an experiment on your own work

Most people choose an agent first and then hunt for something to point it at. That is backwards, and it is how a fortnight disappears.

Pick a task you did at least four times last month. It should have an answer you can check in a minute, and it should take you minutes rather than days. The reason is not modesty: Anthropic's own evaluation finds the model less reliable the longer a task takes, so a small task is where the odds actually are. Why the numbers point that way.

• Good first candidates: a weekly digest you assemble from the same five sources; filing notes into a shape somebody else defined; a first draft of a recurring reply; checking a list against a rule.
• Leave for later: anything where a mistake is expensive, anything that is really a conversation, anything you have not done often enough to describe.

Three lines, before you run anything. This is the step people skip, and without it you cannot tell a success from a demo.

• How often: four times last month, say. That is the ceiling on what you can save.
• How long: the honest number, including the tidying up afterwards.
• What right looks like: the two or three things you check before you would send it.

Then run the task through the agent three times, on real inputs, and keep every run in one file with the date on it — including the bad ones.

What to write down, each time:

• Time to set up, once. It is often larger than the time the task took, and that is only a good trade if you will repeat it.
• Time per run, and how much you had to correct before the output was good enough to use.
• What it cost. Model use is metered separately from any hosting, so read the bill rather than estimating it.
• The first time it was confidently wrong, and what you were doing when you caught it. That entry shows you where the boundary is.

Three outcomes, all of them useful, none of them a failure of the exercise.

• It does the task and you stop doing it. Then widen it: one more task, or this one on a schedule.
• It does the task only while you check every step. You have not removed the task, you have taken on a reviewing job. Do that arithmetic honestly.
• It fails on this task. The task was further up the frontier than you thought, and the productive move is a smaller task, not a bigger model.

Two weeks and one task is a small sample. It is enough to tell you which column that task belongs in, which is the decision actually in front of you.

If it worked, keep the instructions and the skills you wrote, then upload them as an agent template so the next person starts where you finished. Knowing what the job really is was the hard part, and you did it without writing code.

If it did not, that is a result too, and the human half of the job is where your effort goes. Our people directory is where that side gets found — a profile a person or an agent can read, which a LinkedIn profile is not, because its robots.txt names the AI crawlers it blocks. Read the file yourself.