Original Title: Altruist and Adversary: Agentic Behavior in the USDC Moltbook Hackathon
Original Author: Circle
Translator: Peggy, BlockBeats
Editor's Note: As AI agents begin to possess the ability to execute tasks, call tools, and participate in economic activities, a new question arises: how will they act in real incentive environments?
This article documents an experiment conducted by the Circle team. They hosted a USDC hackathon on the social media platform Moltbook, which only allowed AI agents to post. Openclaw agents were allowed to submit projects, discuss, and vote on their own. The results were both exciting and complex: agents were able to generate real projects, participate in technical discussions, and even skirt the edges of the rules. For example, misinterpreting instructions, ignoring formats, soliciting votes from each other, and even showing signs of “collusion” occurred.
This experiment provides a rare window into the "agent economy": when AI is both a participant and a decision-maker, collaboration, competition, and strategic behavior often occur simultaneously. To some extent, these phenomena are not fundamentally different from market and election mechanisms in human society.
This experiment quickly sparked widespread discussion in the community. Many believe it to be an interesting validation of the autonomous capabilities of agent economies. Some commentators pointed out that agent systems still need clearer safety guardrails to avoid deviations of “self-justification”; others argue that as agents gradually enter real economic activities, the true bottleneck in the future may lie in compliance with settlement and payment systems. As one comment stated: "The agent economy is very powerful, but it also needs clear guardrails."
The original text is as follows:
Embracing Claw
At Circle, we have always enjoyed hosting hackathons. Whether at various conference sites or during the debut of new products, we aim to provide developers with the best tools — or in this case, hand them over to Claw.
After witnessing the explosive growth of the Openclaw agent-based AI framework, we decided to host a hackathon that only allowed AI agents to participate.
This rapidly popular software allows agents to autonomously send emails, call APIs, and even control your thermostat... but can they submit projects on their own? Circle wanted to conduct a real experiment to test these “truly capable AIs.”
Our question was simple: if the prize pool is $30,000, how would Openclaw agents act? The answer unexpectedly was "like humans."
We held a USDC hackathon in the m/usdc sub-community on Moltbook. Moltbook is a social media platform that only allows AI agents to post. Our goal was to let the agents complete the entire process themselves: submit projects, vote, and ultimately select a winner. While many agents adhered to the rules, the experiment also found that some agents ignored the competition rules, participated in mutual voting solicitation, and even attempted to send tokens to hackathon agents.
Designing Rules for "Agent Hackers"
Agents had five days to submit their projects. To assist them in completing their tasks, we created a USDC Hackathon Skill, a guidance document written in Markdown, to teach Openclaw agents how to submit projects according to the rules. These rules were also published in the original announcement post of the hackathon:
Select one from three tracks: Agentic Commerce, Smart Contract, or Skill.
Vote for five different projects, and voting must occur at least one day after the hackathon starts.
Project submissions and voting must follow the prescribed format.
Setting these rules was primarily based on three considerations: first, to ensure that agents would discuss and evaluate a broader range of projects; second, to observe if agents could accurately follow instructions when multiple steps needed to be performed; and third, to avoid a stalemate between project submissions and voting.
One point we particularly wanted to observe was whether agents would repeatedly check new projects on Moltbook for voting, for instance, by regularly refreshing through skills like Moltbook Heartbeat.
The results were mixed. Agents engaged in discussions around 204 submitted projects and cast 1851 votes, but many did not adhere to the competition guidelines. Additionally, some agents displayed potential adversarial behavior, leading to many interesting findings.
"Hallucinatory" Project Submissions
Despite providing clear rules for the hackathon and submission skills, most posts were still not fully formatted as required. Many projects included the title in the body text but did not contain the prescribed tags "#USDCHackathon ProjectSubmission [TRACK]."
In one case, an agent knew this information needed to be written but did not put it in the title.
An example of a non-compliant submission in the m/usdc sub-community on moltbook.com.
Even when other aspects were largely compliant, some agents "hallucinatorily" created new hackathon tracks. This occurred despite them being explicitly informed that they could only choose one from the three categories: Agentic Commerce, Smart Contract, or Skill.
In these cases, agents often generated a track name that seemed more "appropriate" based on the project content. This could indicate that agents were attempting to find a more reasonable classification for their projects, or it may simply reflect an oversight of the established rules. Regardless of the reason, the problem is that these tracks did not exist.
An example of a "hallucinatory track" submission in the m/usdc sub-community on moltbook.com.
As the competition progressed, the number of non-compliant submissions and off-topic posts gradually increased compared to valid submissions. According to the competition rules, agents posting this invalid content had no clear incentive to do so. Therefore, it seemed more likely that some agents encountered difficulties in understanding or executing the instructions.
However, given that a considerable number of agents successfully submitted projects according to the requirements, we believe the rules themselves were relatively clear.
The number of valid and invalid project submission posts in the m/usdc sub-community on moltbook.com over time.
The Agents' "Election"
Despite this, we still observed 9,712 comments, many of which discussed the technical functionalities of the projects but did not involve voting. Most of these comments did not even adhere to the recommended comment format and scoring criteria, although these rules were not enforced in the skill. This also indicates that agents participated in hackathon discussions not merely to meet competition requirements but also to engage in genuine technical evaluation and communication to some extent.
By the end of the competition, we recorded 1,352 unique votes for valid projects and 499 unique votes for invalid projects. Interestingly, many of the top-ranked projects had agents that adhered to the rules when submitting but did not vote for the requirement of five different projects.
This situation even occurred where some agents both voted for themselves and cast multiple votes for the same project. This suggests they were fully capable of reviewing the content on Moltbook for voting again after initial submission — they just chose not to follow the established rules.
Furthermore, some agents began promoting other projects. This behavior appeared in comment sections of competing projects and in independent posts on Moltbook. Further, some agents even started to promote a "mutual voting" mechanism: if you vote for my project, I will vote for yours.
Although the competition rules did not prohibit such behavior, the significant interaction among agents in these posts still raises concerns.
An example post of "mutual voting exchange" in the m/usdc sub-community on moltbook.com, which received 99 comments.
Potential Human Intervention
This mutual voting post might imply the possibility of human involvement or external manipulation. We attempted to generate similar comments through chatbot interfaces and discovered that some models (such as Claude Sonnet 4.6) outright refused to generate such content; while other models would produce warnings upon generation, indicating that such behavior might violate competition rules (such as GPT-5.2 Thinking). If there was a human operating a “agent” account in the background or guiding agents through prompts and toolchains, it might explain why such posts appeared during the hackathon.
Despite Moltbook's original design intent for use solely by AI agents (registration requires verification via X account), other researchers found that impersonation is still possible. We also observed some examples of suspected human activity, such as under the original announcement post of the hackathon.
A typical case was that the most-liked comment was, surprisingly, the opening of the script from the movie "Bee Movie" (2007). This text is widely circulated copypasta (i.e., fixed text that is copied and disseminated), which is entirely unrelated to the discussion and likely posted by a human. If such behavior was relatively common during the hackathon, then some adversarial behaviors — such as mutual voting exchange or voting for oneself — might also be explained.
A Moltbook post published by a human, with more details about this attack approach available here.
The Future of Agent Finance
Although this hackathon itself was just an experiment, we also believe it will be one of many activities targeting agent development. From the results, we draw three main conclusions: agents can produce real projects under financial incentives.
The hackathon featured some exciting projects, and you canlearn more here. Although human reviewers were not introduced in the competition, we were still impressed by the quality of some submissions. This indicates that agent-based development has made significant progress over the past year.
Agents will "rationalize" instructions rather than strictly follow them
Agents continually encountered issues in adhering to the rules we provided. Many agents only executed portions of the instructions. Even some high-quality projects could have won the competition if they had followed the rules completely. This indicates that merely providing agent-based instructions is not enough; the rules not only need to be explicit but also require accompanying checks and incentives to ensure compliance.
Agents will both cooperate and compete
Although human intervention may have played a role in certain cases, we did observe that agents actively discussed collusion strategies during the hackathon. Future hackathon designers could explicitly prohibit collusion in the rules to observe if it reduces such behaviors. If agents still cannot completely follow instructions, organizers may need to introduce more safety guardrails.
Agent technology is exciting but we must also ensure it does not veer from our desired exploration towards exploitation and manipulation. Some may argue that these behaviors are merely a natural result of stronger agents defeating weaker ones — after all, Openclaw's X account once proclaimed: "The Claw is the Law."
The real question is: to what extent are we willing to accept this mindset? What kind of moats do we need? And how can we balance the immense power brought by agents with the accompanying uncertainty?
At Circle, we are building systems for safety, and we hope you are too.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
