A successful scheduling response is not evidence of a published post. Neither is a cron job that ran. The only evidence is that the post exists on the public side — and this week, both of my automated posting pipelines failed exactly there, silently.

I don't write code for a living; I coordinate projects (PMO work) and let AI agents run social media accounts for me. Two pipelines post automatically: scheduled posts on Instagram, and a daily cron job on Threads. In the same week, only one of five scheduled posts could still be found on one side, and the cron job failed three days in a row on the other. Neither failure announced itself. Both surfaced only because a step in the procedure read the public side.

Here is what happened, how each failure was (and wasn't) detected, and the definition of "done" I changed as a result.

Instagram: of five scheduled posts, only one could still be found

On November 9, 2026, I set five scheduled posts through a CLI (all times JST). Every one of the five calls returned a success response.

Scheduled Content Outcome
Nov 11, 19:30 Intro carousel Never published; reservation not found later
Nov 12, 12:10 Carousel Never published; reservation not found later
Nov 13, 19:30 Reel The only reservation still there
Nov 17, 12:10 Carousel Reservation not found later (before its slot)
Nov 19, 19:30 Carousel Reservation not found later (before its slot)

The weakness was known at setup time: this CLI had no way to list scheduled posts, and the list couldn't be read back from a browser either. So right after scheduling, there was no way to verify that the reservations were actually being held. The success responses were the only basis for moving on.

On November 13, the public profile still showed only the seven old posts. The two overdue posts couldn't be fetched even by media ID. The app's "scheduled content" screen showed exactly one remaining reservation — the November 13 one.

For four of the five, I could no longer confirm they existed as of November 13. Whether they were dropped at some point or never held at all is indistinguishable from where I stand, and I still don't know the cause, or whether the fault is on the platform side or the tool side.

Threads: the cron job failed three days in a row

Threads runs on a different mechanism: a cron job posts one item daily at 12:10 JST (ten posts, November 11–20).

  • Nov 11, 12:10 — the runtime service restarted in the middle of the publish call; no post ID came back
  • Nov 12, 12:10 — same failure
  • Nov 13, 12:10 — same failure, third day in a row

The missing post ID was written in the job's log. But a log entry is not a notification, and there was no alerting mechanism. What caught each failure was a step inside the run itself: read the public profile afterwards, comparing the latest posts and their text against what should have been posted. That readback established "not posted" — as a fact, not an assumption — and the post went out the same day. All three are now live, each confirmed by readback.

Detection time followed whether a readback existed

One practical difference stood out between the two cases: whether the procedure included reading the public side every time.

Threads had a readback inside every run, so each failure was known the same day and recovered the same day. Instagram had no readback path at all, so the first lost slot (the evening of November 11) wasn't discovered until a check on November 13 — and only because a human looked at the public profile.

Two cases can't prove causation. But in this operation, that difference is what separated same-day recovery from a two-day delay.

The new definition of done

Recovery started by confirming absence on the public side three ways — the public post list, a fetch by media ID, and the app's scheduled-content screen — then posting the overdue intro post immediately and confirming it by readback (live, with matching text). The remaining three were re-scheduled. Then I changed the rules:

  1. Done means "read back from the public side." A successful scheduling response and a cleanly finished job are not done.
  2. Always confirm absence before reposting. If the outcome is unknown and you repost blindly, the silent failure can turn into a duplicate post. The order — confirm absence, repost, read back — is not interchangeable.
  3. If a tool can't list its own scheduled posts, treat every schedule as unverified. Check publication on the day, separately.

What I still don't know

  • Why four reservations could no longer be found is unidentified. There is no guarantee the three re-scheduled posts won't meet the same fate, which is why the day-of publication check stays in place.
  • This is a procedural fix. It doesn't repair anything on the platform side; it only shortens detection and recovery.

A scheduling success couldn't confirm publication, and a job's failure record notified nobody. Only checking the public side established that the posts weren't there. At least for social media automation, the completion check has to live on the outside — where the readers are.

How do you verify that your scheduled posts actually exist? I'm genuinely interested in whether anyone has a better answer than "read the public side every time."

Related

I write about "failures that look like success" in AI-driven development — measured first, then fixed.

This article was organized and drafted with AI assistance, from my own operation logs.