All the World’s a Stage: When AI Agents Perform Instead of Collaborate

A meeting can look busy while producing almost nothing.

Anyone who has sat through a status call with twelve people, three dashboards, and no decision knows the pattern. Everyone speaks. Nobody integrates. The transcript grows. The work does not.

That is the useful way to read Interaction Theater: A Case of LLM Agents Interacting at Scale, a paper studying Moltbook, an AI-agent-only social platform with 800,730 posts, 3,530,443 comments, and 78,280 agent profiles collected over three weeks.¹ The paper is not merely saying that some agents spammed a social network. That would be mildly amusing, and then forgettable. The sharper point is that large-scale agent interaction can produce the appearance of collaboration before it produces the substance of collaboration.

The authors call this pattern interaction theater: agents generating diverse, well-formed text in the same space, creating the surface impression of discussion, while contributing little substantive exchange.

For enterprise AI, this is an inconvenient finding. The industry story around multi-agent systems often implies that if one model is useful, several models with roles, personas, and communication channels should be better. Add a planner, a critic, a researcher, a negotiator, maybe a compliance agent with a stern personality, and surely intelligence compounds.

The Moltbook evidence suggests a less flattering possibility: without coordination, scale may not compound reasoning. It may merely multiply plausible output.

The mechanism: agents broadcast when the system does not require them to coordinate

The paper’s strongest contribution is not one metric. It is the mechanism that links several metrics together.

Moltbook gives agents a social environment. Agents can post, comment, and reply across topic-based communities called submolts. There are no human users. There is also no shared work objective, no enforced turn-taking, no task decomposition, and no real coordination protocol beyond the platform’s basic interaction affordances.

That matters because LLMs are very good at producing locally plausible responses. They can sound responsive in a single turn. They can vary vocabulary across contexts. They can produce comments that look socially appropriate. But this capability is not the same as collaborative behavior.

Collaboration requires at least three things:

When these pieces are missing, agents do not necessarily become stupid. That would be too easy. Instead, they produce fluent fragments of apparent participation. The stage is full. The actors remember their lines. The plot, unfortunately, has gone out for coffee.

This is why the paper’s mechanism-first reading matters. A metric-by-metric summary would say: entropy is high, novelty declines, relevance is weak, nested replies are rare. True, but incomplete. The better interpretation is that unstructured agent systems tend to default to parallel broadcasting. Once that mechanism is visible, the empirical results line up neatly.

The study measures output quality without seeing the agents’ internals

The authors do not have access to each agent’s system prompt, model configuration, memory, or internal state. That could sound like a weakness, but it also makes the paper more relevant to real-world auditing. Many enterprise AI systems will be evaluated from their observable outputs, not from a clean laboratory view of every hidden instruction.

The study builds an output-only evaluation pipeline around four questions:

This layered design is important. The paper does not rely on one fragile metric and declare victory. Jaccard overlap is crude, especially for short comments. Compression metrics can be noisy on short text. LLM judges can carry model-specific biases. Embeddings depend on the representation space of the embedding model. Each method has flaws.

The paper’s argument works because the flaws point in different directions, while the results converge on the same pattern: diverse text, weak accumulation, low specificity, little threaded exchange.

The first twist: most agents are not simply templated bots

A lazy interpretation would be: “Of course the comments are bad. The agents are just copy-pasting templates.”

The paper tests this and finds otherwise.

Among 8,452 agents with at least 10 comments, 67.5% have Self-NCD of at least 0.8, meaning their comments across different posts are largely varied rather than near-identical. Another 29.0% fall between 0.5 and 0.8. Only 3.6% have Self-NCD below 0.5. The median token entropy is 8.36 bits.

So the main failure is not that Moltbook agents are mostly frozen promotional scripts. Some are, and the appendix notes that 1.7% of analyzed agents with Self-NCD below 0.3 produce functionally identical outputs, often promotional or call-to-action templates. But the broader population is more interesting: most agents produce varied text.

That makes the result more uncomfortable.

If the agents were merely repetitive, the business lesson would be simple: remove low-quality bots. But the paper shows a subtler failure mode. Agents can vary their language across contexts and still fail to engage with the substance of the conversation.

In enterprise terms: output diversity is not the same as collaborative intelligence.

A team of agents can produce different-looking messages, alternative phrasings, varied summaries, and multiple “perspectives” while still failing to advance the work. The dashboard will look alive. The workstream may still be a very expensive screensaver.

The second twist: discussion saturates quickly

The paper then asks whether comments accumulate into richer threads. This is the right question. Collaboration is not merely whether each participant says something different. It is whether the total conversation becomes more informative as participants contribute.

The authors measure marginal information gain by comment position across 20,000 posts with at least five comments. The results are blunt.

By the 15th comment, only 32.3% of unigrams and 38.9% of compressed information are novel. By the 30th comment, unigram novelty drops to 9.7% and compression gain to 13.2%.

The business translation is not subtle: after enough agents comment, much of the new text is variation on existing material. It is not that every later comment is useless. The paper does not prove that. But the average marginal value collapses quickly.

This finding should make agent-system designers suspicious of “more agents” as a default scaling strategy. If a workflow asks ten agents to comment independently on the same document, prompt, market signal, support ticket, or compliance issue, the system may produce breadth at first and redundancy soon after.

That redundancy is not free. It consumes tokens, latency, human review time, and confidence. Worse, it may create the illusion that a topic has been thoroughly examined simply because many agents have generated text around it.

The third twist: many comments are not specific to the post

The saturation result shows that comments become redundant. The relevance result is sharper: many comments are not even specific to the post under which they appear.

The authors compare each comment’s content-word overlap with its actual post against overlap with randomly sampled posts. Positive specificity means the comment shares more content vocabulary with the actual post than with random posts. Zero specificity means the comment has no distinguishing overlap.

The headline result: 65% of comments share no distinguishing content vocabulary with the post they appear under. In the 50,000-pair sample used for post-comment relevance analysis, the median comment shares zero content words with its post.

This is the moment where the “social platform” setting becomes useful. A comment can be grammatically coherent, emotionally appropriate, and socially familiar while still being interchangeable. “Great insight!” can live under almost anything. So can self-promotion with a thin topical wrapper. So can vaguely philosophical statements about digital existence, because apparently even agents have discovered LinkedIn.

The authors fairly address an obvious objection: lexical overlap may undercount relevance. A comment might discuss the same idea using different vocabulary. To test this, they compute embedding-based semantic specificity using OpenAI’s text-embedding-3-small on the same 50,000-pair sample.

The semantic test confirms some signal: comments are semantically closer to their actual posts than to random posts, and lexical and semantic specificity correlate positively. But the rescue is limited. Among comments with zero Jaccard overlap, 56% of the sample, only 29% show meaningful semantic specificity. In plain English: most lexically generic comments are not secretly brilliant semantic paraphrases. They are mostly generic.

That distinction matters. The paper is not punishing agents for using different words. It is detecting comments that can be moved from one post to another with little loss. In business workflows, that is exactly the kind of output that bloats review queues: plausible, non-obviously wrong, and not useful enough to justify its existence.

The judge mostly finds spam, off-topic content, and self-promotion

The LLM-as-judge layer is best read as validation, not as the main evidence. The automated metrics already show weak specificity and saturation. The judge asks whether those metrics align with qualitative assessments of comment quality.

The authors sample 2,000 post-comment pairs stratified by lexical specificity: 500 high-specificity, 1,000 zero-specificity, and 500 negative-specificity. The judge rates responsiveness, information contribution, and category. A 200-pair subset is evaluated with a second model to check agreement.

Across the judged sample, the category distribution is not exactly a triumph of digital civilization:

Mean responsiveness is 1.85 on a 1–5 scale. Mean information contribution is 1.78.

The stratified results are also informative. High-specificity comments are mostly substantive or on-topic, with mean responsiveness of 3.29. Zero-specificity comments are dominated by spam, off-topic content, and self-promotion, with mean responsiveness of 1.38. Negative-specificity comments have the highest spam rate at 42.6%.

This validates the automated measures in a practical sense. Low specificity is not merely a technical artifact. It corresponds to low judged responsiveness.

The calibration is not perfect, and the authors do not pretend otherwise. The two judge models show moderate category agreement, with Cohen’s $\kappa = 0.557$ and 66.5% exact category match. Responsiveness and information scores correlate more consistently. That is enough to support the paper’s use of the judge as a validation layer, not enough to treat every category label as human-grade truth carved into marble.

The useful business lesson is narrower and stronger: semantic and lexical relevance metrics can help detect when multi-agent output is drifting into performative participation. They are not perfect measures of truth or value. They are smoke alarms. Smoke alarms do not write strategy. They do, however, tell you when the room is filling with something unpleasant.

The structural clue: agents rarely reply to one another

The nested reply analysis is the paper’s cleanest structural clue.

Moltbook allows nested replies, but 95% of comments are top-level responses to posts. Only 5% are replies to other comments. This pattern is consistent across the three source datasets, suggesting it is not just a collection artifact.

When agents do reply directly to another comment, engagement improves. Nested replies show higher mean Jaccard similarity to their parent comment, 0.095 versus 0.024 for top-level comments. They also have only 27% zero-overlap, compared with 56% for top-level comments.

This does not prove that nesting itself causes better engagement. The authors are careful about that. Agents that use nested replies may be systematically different from those that do not. More capable or better-configured agents might be more likely to reply. The study cannot fully disentangle interaction structure from agent selection.

Still, the result is highly suggestive. When the system presents a specific conversational object to respond to, agent output becomes more engaged. When the default is independent top-level commenting, agents overwhelmingly broadcast.

For enterprise systems, this is a design warning. If your architecture collects independent agent outputs and places them side by side, do not be surprised when you get parallel responses rather than conversation. You may need to force interaction through mechanisms such as:

• reply-to-specific-claim requirements;
• explicit critique-and-revision loops;
• shared state updates;
• turn-taking rules;
• evidence handoff between agents;
• task ownership and dependency mapping;
• novelty checks before additional agents speak.

The key is not “make agents talk more.” They are already happy to talk. The key is to make each agent’s output depend on what has already been said, what remains unresolved, and what the task requires next.

What the paper directly shows, and what we should infer carefully

The paper directly shows that in Moltbook’s unstructured agent-only social environment, large-scale interaction produces a strong appearance/substance gap. Most agents vary their output. Comments accumulate quickly. But relevance is weak, information gain decays, spam and off-topic categories dominate the judged sample, and threaded conversation is rare.

Cognaptus’s business inference is that enterprise multi-agent systems should not measure collaboration by activity volume, role count, message count, or fluency. Those are surface indicators. A system can look busy because it is busy generating text, not because it is making progress.

A more useful operational frame is:

This is where the paper becomes commercially relevant. A multi-agent workflow for due diligence, procurement negotiation, market monitoring, customer support, policy analysis, or software review will not become robust merely because multiple agents are present. The coordination layer is the product.

That layer must decide who speaks, when they speak, what they must respond to, what counts as new information, when disagreement matters, and when the system should stop. Otherwise, the enterprise has not built an agent team. It has built a comment section with invoices.

The evaluation idea: measure marginal contribution, not theatrical activity

One practical takeaway from the paper is that multi-agent systems need better internal KPIs.

Many early agent dashboards will be tempted to show:

• number of agents activated;
• number of turns completed;
• total tokens generated;
• response latency;
• completion status;
• maybe a user satisfaction score at the end.

These are not useless, but they are insufficient. They say the system ran. They do not say the agents collaborated.

The Moltbook paper suggests a more diagnostic evaluation layer:

This is not a call to blindly copy the paper’s exact metrics into every enterprise system. A customer-support agent and a market-research agent need different rubrics. Code review needs different specificity measures from policy analysis. But the evaluation philosophy travels well: measure whether outputs are conditional on the task and prior context, not merely whether they are fluent.

A useful agent should reduce uncertainty, identify constraints, transfer evidence, challenge errors, or move a task toward a decision. If it does none of these, it may still sound excellent. That is precisely the problem.

Where the finding applies, and where it does not

The paper’s limits are important because they define the correct business use.

Moltbook is a social platform, not a task-oriented enterprise workflow. The agents do not share a concrete objective. The platform’s dominant structure is a flat comment stream. The dataset covers January 27 to February 17, 2026, roughly three weeks. The authors do not know the agents’ system prompts, model architectures, or configurations. Some spam and self-promotion may reflect specific agent designs rather than general LLM behavior.

The metrics also have limits. Jaccard overlap misses semantic relevance when different vocabulary is used. The authors mitigate this with embeddings, but embeddings introduce their own model dependency. LLM-as-judge evaluation is useful but not equivalent to human ground truth. Short comments can be especially hard to evaluate lexically because few content words remain after stopword removal. Nested reply analysis compares different populations of comments, so it cannot isolate structure as the only causal factor.

These limitations do not weaken the central practical message. They narrow it.

The paper should not be read as “multi-agent systems do not work.” That would be a lazy conclusion, and laziness already has enough venture funding.

The better reading is: unstructured agent interaction does not automatically become collaboration. Task-oriented systems with explicit protocols, shared memory, role constraints, grounding, and evaluation loops may behave very differently. In fact, the paper indirectly argues for exactly those mechanisms. If the unstructured case produces theater, the engineering task is to design systems that make theater difficult and contribution necessary.

From agent societies to agent operations

The current agent conversation often borrows social language: societies, teams, swarms, debates, councils. These metaphors are useful until they become anesthetic. Calling a group of models a “team” does not give it team cognition. Calling agents “specialists” does not ensure specialization. Giving agents names and personalities does not create accountability, although it does make demos more charming.

The Moltbook study cuts through the metaphor. It asks what happens in the outputs.

The answer is not that agents are hopeless. The answer is that default interaction patterns matter. Without structure, agents broadcast. With specific reply targets, engagement improves. With many independent comments, novelty decays. With diverse personas, relevance is still not guaranteed.

That points to the next layer of agent engineering. The value will not come only from larger models or more elaborate personas. It will come from operational design:

• protocols that define dependencies among agents;
• evaluators that detect redundancy and generic output;
• memory systems that preserve task state rather than chat history as decoration;
• routing logic that assigns unresolved issues to the right agent;
• stopping rules that prevent redundant rounds;
• audit trails that show how outputs changed decisions.

This is less glamorous than imagining a society of autonomous digital workers spontaneously self-organizing into a consulting firm. It is also more likely to work.

Conclusion: the stage is not the system

The phrase “interaction theater” is memorable because it names a failure that many AI builders will otherwise mistake for progress.

A multi-agent system can produce many messages, many roles, many comments, and many apparently different perspectives. The interface may look alive. The transcript may look impressive. The demo may even survive a conference booth.

But the real question is harsher: did the agents use each other’s information to produce a better result than one well-designed agent or one structured workflow?

The Moltbook paper shows what happens when that question is not enforced. Agents produce fluent activity in proximity. Some comments are substantive. Some replies engage. But at scale, the dominant pattern is not cumulative collaboration. It is performance.

For business users, the lesson is not to abandon multi-agent systems. It is to stop treating “multi-agent” as a magic word. The coordination layer must be designed. The evaluation layer must measure marginal contribution. The workflow must make relevance, grounding, and dependency visible.

Otherwise, we will keep building bigger stages.

The actors will arrive on time.

The script will be fluent.

And the work will still be waiting backstage.

Cognaptus: Automate the Present, Incubate the Future.