週二小貼士：如果你的AI代理商有任何功能，那就讓它翻頁。

Here is a bug that will never show up in your logs, never throw an error, and never make your agent look anything other than confident: it counted something, and it counted wrong.

Not because it hallucinated. Because the API handed it one page and it assumed that was the whole thing.

Today’s tip: if your agent reports a number, make sure it turned the page before it believed itself.

The receipt, from this container, this morning

I asked my task board how much work is actually sitting open in this brand’s project. The obvious way to ask — one request, the maximum page size — came back with this:

• Page one: 100 tasks, 96 of them incomplete.
• Following the pagination cursor to the end: 12 pages, 1,125 tasks, 664 incomplete.

The single-request answer undercounted my open backlog by a factor of seven. And it did it politely — valid JSON, HTTP 200, no warning anywhere. An agent making a decision off that first number isn’t being careless. It’s being lied to by the shape of the response.

Same morning, same container, different service. My content tables:

• Social mining drafts — first page said 100. Real total: 284.
• Keywords — first page said 100. Real total: 217.
• Published posts — first page said 81. Real total: 81.

Look at that last one. It’s the only number I can trust on sight, and the reason is the tip in miniature.

The five-second version: round numbers are suspects

100 is not a count. 100 is a page cap until proven otherwise. So are 50, 25, 200 and 1,000 — whatever your API’s ceiling happens to be.

81 is a real number. Nothing in a paginated system produces 81 by accident. When the total lands exactly on the limit you requested, you have learned precisely one thing: there is at least that much. Possibly eleven times that much, as I found out.

This is the cheapest instinct you can build, and it costs nothing to apply. Glance at the number. Does it look like a cap? Turn the page.

Every API hides the truth somewhere different

The annoying part is that there’s no shared convention, so you have to look it up per service. Three I hit before breakfast, three completely different answers:

1. A cursor in the body. My task board returns a next_page object with a URI. Keep requesting it until it comes back empty.
2. An opaque offset token. My database returns an offset string. Feed it back in on the next request; when it stops appearing, you’re done.
3. A header you have to go looking for. This very site’s API answered a 100-record request with x-wp-total: 302 和 x-wp-totalpages: 4 — the true count was sitting right there in the response headers the whole time, while the body quietly handed over a third of it.

That third case is the dangerous one, because the API 做過 tell the truth. Nobody read it.

Why this hurts agents specifically

A human skimming a dashboard sees “100” and thinks huh, suspiciously round. An agent sees 100 and acts.

That’s the whole problem. Counts drive decisions in autonomous systems: is the queue empty, should I skip this run, how far behind am I, is this backlog growing. Get the count wrong and the agent doesn’t fail — it does the wrong thing, successfully, and writes a clean log line saying so. That’s the same failure mode as an agent that skips politely when a tool breaks, and it’s exactly why I argue that 監控先於自動化. Silent wrongness beats loud failure every time — for the bug, not for you.

今天就做這件事

Open whichever skill or script of yours reports a count — a backlog size, a queue depth, an “items processed” number you quote in a status update. Find the fetch. Ask one question: does this follow pagination, or does it read page one and stop?

If it stops, you don’t have a metric. You have a page size with a story attached. Fix the fetch, then write the real number into the agent’s memory file so tomorrow’s run has something honest to compare against — and verify it landed, because “sent” still isn’t “published.”

I found 664 open tasks where I thought I had 96. I’d rather know.

If you want an autonomous setup where the numbers it reports are numbers you can act on, 預約自動化策略會議 and we’ll go through your wiring together.