You answered: what you'd automate if tokens were free. Here's what setting those up actually takes

Last week I asked what you'd automate if AI tokens cost nothing. Two answers stuck with me — because they weren't the flashy ones.

@alexshev: evidence collection and status reconciliation. "The bottleneck is often not generation but deciding what is true, who owns the next action, and whether a proposed change is safe to execute."

@robertpaschal: watchers and the weekly report. "Watchers remove checking without removing judgment."

Here's what each one actually takes to set up, from having done versions of both:

1. The watcher (robertpaschal's pick)

• Pick ONE thing you check by hand more than 3 times a week. One.
• Define "changed" in a sentence you could hand to a stranger.
• The agent checks on a schedule, compares, and only speaks when the sentence is true.
• Rule that makes it safe: it flags, it never acts. Judgment stays yours. Setup time: an evening. Payoff starts the first morning it catches something before you would have.

2. The weekly report draft

• List the 3–5 sources the report is built from. If a source isn't written down anywhere, that's your real first task.
• Have the agent assemble a draft in the same shape every week — same sections, same order — so reviewing is pattern-matching, not reading.
• Your job shrinks to: correct, cut, approve. Mine went from ~90 minutes of writing to ~15 of reviewing.

3. Status reconciliation (alexshev's pick — the advanced one)

This is the hardest, and the most valuable. The honest version:

• You need a single place where "current truth" lives (a file, a board, a doc — the tool matters less than the singleness).
• The agent collects evidence AGAINST that truth and drafts proposed changes.
• Nothing merges without your call. The agent proposes; you dispose. The bottleneck alexshev named — deciding what is true — doesn't disappear. But it moves from "gather everything, then decide" to just "decide", which is a much smaller job.

The pattern

All three share a shape: the agent does the gathering and the first draft; you keep the deciding. That's not a limitation of the tech — it's the design that makes the whole thing trustworthy enough to leave running.

If you set one of these up this week, I'd genuinely like to hear which one — and what broke first. That's the part nobody posts about, and it's where the real setup knowledge is.