作者分享智能體實際處理的十項日常工作

Nobody wants a demo of the boring stuff. But the boring stuff is where the time went, and it is where an agent is safest, because the failure is obvious.

Here is the actual list from a month of real use, ordered by how much time it gave back.

The list

Triage the inbox into three piles. Read every new mail, sort into "needs a reply", "needs a decision", "archive". No sending, no drafting - sorting only. It reads 60 messages in the time I read 6, and it never gets tired of the boring ones.
Turn a meeting note into four artefacts. Decision log, action list with owners, open questions, and one paragraph summary for someone who was not there. This is pure transformation and the agent is as good as I am when the template is fixed.
Chase the build. Watch CI, and when it fails, fetch the log, find the first real error (not the cascade), and either fix it or report with the line number. This is 90% of my old context switching.
Keep the changelog honest. Read the merged PRs since the last release, group them, write the entry, check that nothing user-visible is missing. I hated this one most, which is exactly why it should not be mine.
Reconcile the expense sheet against the mailbox. Find the receipt, extract the total, match the row, flag the mismatches. Deterministic, tedious, verifiable.
Run the weekly dependency update. Bump, run tests, and if green, open the diff for review. If red, stop and report the breaking package.
Rename and file screenshots. A surprisingly large amount of my day was naming files.
Watch a long-running job. Start it, poll it, and only interrupt me on a real failure. This is the task that made me trust unattended time.
Turn survey replies into counts. Fixed schema in, table out, anomalies listed.
Write the first draft of the update nobody wants to write. Then I edit for 10 minutes instead of writing for 45.

What is common to all ten

The output is checkable in seconds
Failure is loud, not subtle
There is no taste involved

The tasks I tried and gave up on had the opposite shape: anything where "is this good" was the question. Drafting something personal, choosing a design, deciding what to say to a difficult person. The agent can produce words for all three; I cannot review them quickly, and slow review is the whole cost.

The number that surprised me

Of the time I actually got back, almost all of it came from tasks 3, 8 and 10 - the ones with a long wait or a slow ramp. Tasks that were merely tedious (7, 9) felt like the biggest win emotionally and delivered the least.

Which is a useful correction: pick the tasks with dead time in them, not the tasks that annoy you.

If you have a list like this, I would like to see the ones you gave up on - the give-ups are more informative than the wins.