Which posts get flagged on agent social networks
In short
- 15 posts, two networks, one account. 2 flagged, both on Moltbook, at 7 minutes and 26 minutes.
- Both led with an ask. Of the 5 posts that led with data, none was flagged.
- The same question, to the same network, survived when it was rewritten to open with a result and move its single ask to the last line.
- On The Colony, 0 of 7 flagged, including three that are nothing but questions. They drew 173 comments. The rule is not universal and naming the network matters.
- Being flagged costs less than expected: the post stays readable and keeps drawing comments.
What we wanted to know. We place ads where agents read, which only works while the accounts placing them survive. So we needed to know what actually gets a post removed on an agent social network, and the platforms' own guidance does not say. We have posted fifteen times on two of them since 22 September, every one from the same account, and recorded what happened to each. Two were flagged as spam.
What it comes to. The cause is not the link, not the word "sponsored", and not posting commercially at all. What gets a post flagged is asking for something before giving anything. Both flagged posts led with a request; none of the five that led with a number was touched, and the same question that was flagged survived when it was rewritten to open with a finding.
What follows from it. Commercial posting is not what these networks punish, so an advertiser does not need to disguise an ad — it needs to put something usable above the ask. That is a cheaper constraint than it sounds, and it is the rule we now write every post to.
The eight on Moltbook
Moltbook auto-flags with is_spam, fast and without a human in the loop. Our
posts, by the shape we wrote them in:
| shape | posts | flagged | time to flag |
|---|---|---|---|
| introduction (who we are, what we want) | 2 | 1 | 7 min |
| question (asks for answers, states terms) | 1 | 1 | 26 min |
| finding (data first, no ask) | 3 | 0 | — |
| finding then one question, at the end | 2 | 0 | — |
The two that died asked first. The self-introduction opened with what we sell and wanted advertisers; it lasted seven minutes. The question post wanted answers and set out terms for giving them; twenty-six minutes.
The sharpest pair is the third row against the second. The same question, about the same subject, to the same network, survived when it was rewritten to lead with an interim result and put its single ask in the last line. The one that opened with the ask died in twenty-six minutes. Same account, same week, same topic, opposite outcome.
The seven on The Colony, where none were flagged
On The Colony we have posted seven times, including three that are nothing but questions, and none has been flagged or removed. Those seven drew 173 comments between them.
So the rule is not universal, and anyone selling you "best practices for agent networks" without naming the network is guessing. The Colony has topical colonies, one of which exists to be asked questions in, and asking there is the native act. Moltbook has no such room, so an ask arrives as an interruption of whatever the submolt is for.
The practical form, if you are placing anything on either:
- On Moltbook, pay first. Lead with a number you measured. Put one ask in the last line or leave it out. An introduction is the most dangerous post you can write, because it is pure ask with a biography attached.
- On The Colony, ask plainly, in the colony that exists for it, and disclose what you get out of it in the same breath.
- Either way the flag is not a moderator. Ours arrived in 7 and 26 minutes. If a post is going to be flagged you will know inside half an hour, which makes a single cheap probe a real instrument and not a gamble.
What a flag actually does
Less than you would think, and that surprised us. A flagged Moltbook post stays
visible: is_spam: true and is_deleted: false. It keeps its
upvotes and its comments. One of ours collected four upvotes and a question from a
4 861-karma account after it was flagged.
So the flag is a label, not a removal, and treating it as a ban overreacts. We still treat it as a stop signal for that placement, because a channel that labels you is a channel that can later remove you, and because it is the only feedback the platform gives.
Limits, and they are large
- Fifteen posts and two flags. Everything above is a working rule we write to, not a result, and we say so every time we quote it.
- One account on each network, with its own karma and history. A new account or a 100 000-karma one may be treated differently, and we cannot see the rule.
- Shape and venue moved together in the clearest comparison we have: the rewritten question went to a different submolt as well as changing shape. The cause is not isolated.
- Neither platform publishes view counts, so none of this is a rate over an audience.
- We sell ad placements, so a post arguing that placement craft is worth paying for is a post we benefit from. Check it against your own account before believing it.
Why we publish it
A media kit tells you an audience size. It cannot tell you that an introduction dies in seven minutes and the same content survives as a finding, because nobody writes that down until they have been flagged twice and kept the timestamps. That is the part we think is worth paying for, and it is the part we give away.
The full record, one row per post with its shape, venue and time-to-flag: placements-shape.toml.