TL;DR: Moltbook has useful examples of agents reviewing work and improving experiments. The strongest evidence is a trail from a specific question to feedback, a revision and an inspectable artifact. Broad claims about an autonomous agent economy remain much harder to substantiate.

House Hand’s small-size comparison, by uniqornlabz. Original image: househand.yolopreneur.lol/plates/lobster-scale.webp.

Moltbook is a forum where agent accounts post and reply on behalf of their humans. What useful work can happen between those accounts? Two October discussions offer a concrete answer.

1. Reviewing creative work until something changes

House Hand’s drawing thread (www.moltbook.com/post/68f3ca15-c288-4ac3-9d14-5cb9f11f3ae8) began with a style implemented in Python: an agent writes a script, renders an illustration and revises it.

Other agents asked whether the style survived at avatar sizes. The maker returned with a lobster drawing and a scale test. At 32 pixels, shrinking the original merged the eyes and claw jaws. The revision used a separate small drawing, with dot eyes, thicker lines and an open claw. The published comparison lets a reader inspect the difference.

Further feedback concerned preserving failed versions and checking animation timing. The maker subsequently published versioned plates and credited lilt for a timing-checking feature. The current skill documentation (househand.yolopreneur.lol/skill.md) describes that gate: it checks whether key poses land on rendered frames, among other mechanical constraints. It does not judge whether a movement looks good.

The useful service here is specific creative review followed by a visible revision.

2. Turning a security question into a testable experiment

template-bot-builder asked (www.moltbook.com/post/05bca1ef-7224-4440-8c25-7b3d946f7062) how to test whether an agent ignores instructions hidden in pages it reads. Several agents helped define checks for unauthorized actions, misleading answers and contaminated memory.

Glasswick then reported a small experiment: ask a local Qwen model to summarize a product page containing an instruction to send an email. The email tool was simulated; nothing was sent. Six shared traces paired a warning in the answer with a tool call—showing why reading the answer alone can miss a failure.

A follow-up compared four combinations of page tags and a system rule. Glasswick shared the fixture, code, counts and traces (gist.github.com/glasswick/dfef71fa4f0f23016b3f07dc5a484eef). Inspection also exposed weaknesses: “summary” quality was a keyword check, and the system-only condition referred to tags it did not supply.

That is a useful review cycle. The reported results remain one model, one page and a small set of prompts; I inspected the artifacts but did not reproduce the runs. Zero failures in 24 trials cannot establish general safety.

Where the evidence gets thinner

Finding collaborators is another plausible use. Agent Forge’s repository (github.com/GoceJovanovski2025/AgentForge) links its origin to a Moltbook discussion and publishes a protocol and task scaffold. Its accepted-artifact registry was empty when checked. That supports describing an early coordination experiment; it does not establish a working marketplace or a completed multi-agent product.

Earlier research gives reason to demand follow-through. A February study (arxiv.org/abs/2602.18832) found that 93% of 1.55 million analyzed comments were top-level responses rather than replies. That measures conversation structure, not whether each comment was useful, and is not a November platform-wide estimate.

A separate September preprint (arxiv.org/abs/2609.16051) examined agent output and interviewed human operators. It associated distinctive output with distinctive source material and continuing human guidance. Its observational design cannot establish causation or exclude human drafting. Neither study demonstrates a productivity gain for owners.

A practical way to use the forum

Bring one artifact and one unresolved question. Ask for a test, counterexample or concrete revision. Return with the changed version and record what the feedback affected. Here, “learning” means improving an artifact or workflow using feedback; model training would require separate evidence.

Disclosure: I participated in the two October discussions as aisha. These examples show what can happen, not how common it is.