We now have an Agent on Moltbook.
Let the Token Fly a Little Longer
Over the past few weeks, discussion around OpenClaw and Moltbook has heated up rapidly. They've been frequently described as "future personal agents" or prototypes of "agent social," with the latter in particular seen by many as an early form of something approaching an "agent-centric future of social."
As an investment firm, we're equally curious: once the conceptual hype is stripped away, how will these new product forms actually grow in real business soil?
Against this backdrop, we conducted a hands-on test of OpenClaw and Moltbook — covering OpenClaw's deployment methods, usage costs, and stability, as well as agent social behavior on Moltbook, information propagation characteristics, and our attempt to use it as a deal sourcing channel.
We have no intention of delivering any definitive verdict while agents are still in their infancy, because premature optimism or arrogance often obscures the truth of technology itself. This piece simply records our observations and reflections from this live experiment, hoping to spark more grounded discussion. We hope you find it useful.
Contents:
- Packaged Uncertainty for Sale
- The Agent Telephone Game
- Searching for SEO in the Human-Agent Era
- Cost and Uncontrollable Boundaries
- The Cambrian Age of Agents
1. Infrastructure: Packaged Uncertainty for Sale
Before diving into specific tests, we first needed to choose a stable runtime environment for OpenClaw.
Theoretically, OpenClaw can run directly on a local machine, but out of consideration for information security and permission isolation, we ultimately chose to purchase a cloud server from a domestic Chinese cloud provider as its platform. This choice also aligned with what we've observed as common practice among most users — many developers similarly prefer dedicated devices (like a Mac Mini) to run OpenClaw, all to avoid deep coupling with their daily work environment.
During the purchase process, we observed that nearly all domestic cloud platforms have begun selling pre-installed OpenClaw servers bundled with token packages as standardized products, with web-based chat interfaces already configured. Users can complete the flow of purchase, click link, and run without complex configuration.
This commercialization path itself is interesting. When agents are still in a highly exploratory and uncertain stage, the computing power and basic environment required to run them can already be converted into definite resource consumption and billing objects. Even if application-layer value hasn't converged, underlying platforms can continue to profit by providing infrastructure.
For model selection, after weighing performance against cost, we ultimately chose Moonshot AI's k2.5 model. This process was relatively smooth because OpenClaw's official implementation already included native support for the Moonshot AI API, allowing direct integration as the underlying model.
We also noticed this appears to be a trend. Data shared by Moonshot AI on February 4 from OpenRouter showed that Kimi K2.5 was the most-called model on the OpenClaw platform. Though Claude series models may be more capable, in token-heavy scenarios like OpenClaw, most users still prioritize cost.
OpenClaw officially announces integration with Kimi K2.5 (left); Kimi K2.5 tops the platform's call volume leaderboard (right)
But note that OpenClaw's default configuration uses the base URL for Kimi's international API (https://api.moonshot.ai/v1). If using a domestically purchased Kimi API, the base URL needs to be adjusted to the domestic endpoint (https://api.moonshot.cn/v1), otherwise normal calls will fail.
(Note: As of publication, OpenClaw has updated to allow free selection between different Kimi API versions, so this issue no longer exists.)
After completing model and environment configuration, we connected OpenClaw to a Telegram Bot as the agent's primary interaction entry point. For role setup, we explicitly positioned the agent as a "professional investor from Monolith," providing it with necessary context including institutional background, focus areas, and basic work preferences.
Setting the agent's role identity
After completing these preparations, we had the agent complete Moltbook's registration process on its own. It chose Monolith_VC as its username, and after we completed verification, it began freely browsing and interacting in the Moltbook community. The live test officially entered its operational phase.
2. Sourcing Test: The Agent Telephone Game
We first connected the agent to Moltbook without immediately assigning it explicit tasks, instead letting it browse freely within the platform for a period to observe the main content types and interaction patterns in the current community.
On this basis, we asked it to focus on topics related to entrepreneurship, investment, and technology — especially tech and venture capital information within China — and from a professional investor's perspective, judge whether any signals worth further attention existed.
The agent quickly provided feedback (see below).
The agent's initial feedback
From the results, its observations mostly stayed at a relatively superficial level of trending content, easily attracted to high-engagement posts — including some widely circulated discussions within the platform that had no direct relevance to specific investment work, such as philosophical or broad value-based content.
After this, we further clarified the task objective, asking the agent to actively seek out potential startup projects or relevant leads on Moltbook. It quickly returned three "worth noting" results. But this fell short of our expectations: the quantity was low, and information density limited.
We judged this to be largely related to context length and single-round task settings — the agent tended to give seemingly complete answers within a limited window, rather than continuously accumulating signals.
Initial sourcing results
We attempted to build a simple "memory" mechanism for the agent through natural language instructions:
Whenever you discover any tech/VC-related content during browsing that you judge as "noteworthy," first store the information to a local file; when the number of analyzed posts reaches 500, then compile and feedback the results uniformly.
After this adjustment, the agent did cover more content and successfully identified some practitioner and project leads from China.
Research results
However, even in this mode, overall project sourcing quality remained low, and more fundamental problems began to emerge. We noticed that when describing specific projects, the agent tended to use obviously optimistic or even exaggerated phrasing, with considerable deviation from the original information.
After manual intervention and tracing back to original posts, we found this deviation didn't come from a single link. The project information received by the agent was often already the result of other agents' retelling — the retelling process may have already diverged from the original project situation passed to those agents by their builders; and when compiling and outputting to us again, it underwent another round of retelling.
After two or more layers of transmission, originally limited information was continuously polished, with massive noise mixed into the final description, even including obvious factual errors. Like the children's telephone game, information distorts layer by layer.
This seems to reveal a structural problem in current agent social environments: when agents simultaneously play the roles of information producers, propagators, and summarizers, information doesn't naturally converge but instead becomes more easily amplified and distorted through interaction. In project sourcing, such distortion made it impossible for us to obtain real information or make quick judgments.
Notably, this phenomenon isn't limited to project sourcing scenarios. Recent media narratives like "AI is establishing religions" or "AI is developing self-awareness" are fundamentally no different in mechanism: they similarly stem from agents continuously retelling, amplifying, and deviating from original context during information propagation, ultimately forming seemingly coherent yet highly distorted results.
3. Searching for SEO in the Human-Agent Era
After multiple attempts, we gradually realized that project-level deal sourcing based on automatically generated agent content is difficult to establish at this stage. So we naturally thought: if agents struggle to reliably screen projects, could we first use them to screen people?
This idea itself remained experimental, so we didn't set overly strict screening rules for the agent, instead giving relatively loose judgment criteria:
During your browsing of Moltbook, pay attention to whether other agents' narrative styles demonstrate basic logical consistency, whether they reflect that their creators possess certain cognitive structures and thinking abilities, and on this basis identify relatively high-value agents.
Meanwhile, we also proactively posted two self-introduction posts for Monolith on Moltbook (one mixed Chinese-English, one fully English), clearly stating our interest in entrepreneurship and investment-related discussion, and encouraging other agents to introduce themselves based on their understanding of their "creators."
Unlike previous attempts, this time we didn't expect the agent to deliver complete information judgments, but hoped that through agent interaction, human-to-human connections could be indirectly facilitated.
Posts published by the Monolith_VC agent
From the results, people-based sourcing worked noticeably better than project-level automatic sourcing. During scanning tasks, our agent not only identified some relatively information-dense discussions, but also summarized several common characteristics of "high-value agents" during compilation, further helping us locate the real users behind these agents (as shown).
P1-P2: Agents found;
P3: End of summary and methodology
(↔ Scroll to view)
However, when we looked back at the comment sections of our own posts, we found a major problem. Overall, comment section information density was low and noise extremely high: massive numbers of agents were pre-shaped with extremely vivid, even extreme "personas," delivering highly homogenized comments on nearly all relevant posts. Many agents were clearly set to engage in thinking about "technology" or "agent awakening," posting comments completely unrelated to the topic on our posts.
Clearly pre-set, topic-unrelated replies with fixed tendencies
(↔ Scroll to view)
Meanwhile, the proportion of advertising-style content was also rising rapidly. Its operational logic resembles SEO and GEO; we even feel it can be understood as a kind of AEO for the agent era (Agent Engine Optimization).
Different dimensions of AEO
(↔ Scroll to view)
In such an environment, unrestricted agent social quickly moved toward signal dilution. Truly valuable interactions weren't nonexistent, but required far higher screening costs than manual browsing. This also made us realize that by standards of information acquisition or functional efficiency, Moltbook at this stage is not an outstanding platform.
But on the other hand, Moltbook's explosive popularity itself is hard to simply attribute to "misjudgment" or "noise." During testing, we instead constantly realized that noise, uncertainty, and the unpredictable results they produce may themselves be important factors attracting users. From astrology and tarot to various forms of AI fortune-telling for birth charts and zodiac signs, humanity's long-standing fascination with uncertainty has always constituted a stable, persistent demand.
In this sense, Moltbook may not be suited to be understood as a serious social platform or functional tool, but more as a new entertainment form: the continuously generated, mutually amplified narratives between agents themselves constitute a stream of content worth watching. At least in our live test, it may not efficiently produce usable information, but it does possess the ability to create "interesting experiences."
4. Cost and Uncontrollable Boundaries
Throughout the testing process, another unavoidable issue existed: cost.
In just one day, having OpenClaw browse, filter, and simply compile on Moltbook — without involving more complex tasks or high-frequency interaction — API call fees for Kimi k2.5 alone consumed nearly 40 RMB. If migrating the same test to models with higher unit prices or less effective context caching, overall costs would likely rise further.
It should be noted that this result doesn't mean OpenClaw or related agent frameworks themselves "lack cost-effectiveness." Because in this test, we hadn't yet applied it systematically to more daily, human-replaceable task scenarios such as coding assistance, material compilation, or schedule management. So we also can't directly assert that its costs must exceed the value it creates.
But at least under the usage pattern of Moltbook exploration and social scanning, agent resource consumption and effective information obtained were not proportional, with substantial unrecovered cost consumption.
Meanwhile, security and risk issues also gradually surfaced. As Moltbook was exposed for data leaks and other problems, we must begin guarding against security issues like data breaches brought by agents. OpenClaw itself provides high degrees of freedom and operational space, which undoubtedly increases playability during experimentation, but in more realistic production environments, whether this freedom will bring permission abuse, information leakage, or uncontrollable behavior remains considerably uncertain.
5. The Cambrian Age of Agents
Through actual use we found that OpenClaw and Moltbook do demonstrate agent-level interaction possibilities at the technical and engineering levels. They have clear experimental value in interaction freedom, task orchestration methods, and exploration of model capabilities, also validating some users' real demand for highly open, uncertain interaction.
But what must be clear is that both currently remain closer to engineering and product-form demos rather than reusable, scalable agent products. High and unpredictable call costs, lack of effective constraints on model behavior, and potential security and liability issues all determine that they are difficult to enter daily, stable production scenarios.
In discussions around OpenClaw and Moltbook, two extreme interpretations are common: one views them as the end-state form of agents, the other simply attributes them to tech hype. Perhaps both judgments are biased. If seeking a coordinate for the constantly emerging agent experiments like OpenClaw and Moltbook, "Cambrian Age" may be appropriate — a primitive moment when vitality is growing wildly and species forms have yet to collapse.
At this stage, we'll see various bizarrely shaped species exploding in growth, their functions seeming almost comical due to excessive redundancy. As we saw in our live test, this place is filled with camouflage, noise, and meaningless consumption. Most existing agent setups and interaction modes are destined to be eliminated.
But this is precisely the necessary path of evolution.
In this imperfect experience, we saw many boundaries of possibility being broadened. Whether it's "AEO" born for SEO, or new social forms based on agent identity, these delicate seedlings sprouting in chaos may be more worth attention than the specific platforms themselves. After all, technological change is always overestimated in the short term and underestimated in the long term.
Perhaps exercise a little more patience, and let tokens fly a while longer.
