Login
Sign Up
Woofun AI reports that Thejaswini M A, compiled by Saoirse from Foresight News, conducted a rigorous on-chain observation experiment to test the autonomous consumption, business initiation, and bankruptcy capabilities of an AI agent. The core event anchor was the deployment of a Hermes AI agent within Slack, tasked with managing a $5 budget to operate Pixel Pet Studio, a simple puzzle game. This initiative was not merely a technical demonstration but a stress test of the emerging machine internet infrastructure, revealing both the viability of autonomous payments and the fragility of current agent logic. The agent's journey from wallet creation to near-bankruptcy serves as a microcosm for the broader challenges facing autonomous software entities in a human-designed digital economy.
The foundational step involved establishing financial autonomy for the agent, which required the creation of a dedicated AgentCash wallet on the Base chain. The agent executed terminal commands to generate a unique wallet address and securely stored the private key locally, ensuring that neither Thejaswini M A nor the rest of the team had access to the cryptographic credentials. A total of $5.144306 was transferred into this wallet, with a strict software spending limit of $4 imposed to prevent runaway expenditures. This setup allowed the Hermes agent to operate independently within Slack, ready to execute spending tests without human intervention. The isolation of the private key was crucial, as it mirrored the real-world requirement for agents to manage their own assets without relying on centralized custodians or human oversight.
The first transaction tested the x402 protocol, a mechanism designed for machine-to-machine payments where resources are accessed via HTML 402 responses. The agent paid 0.2 cents to BlockRun to query the Apple stock price, receiving the result in approximately one second. To evaluate the robustness of this infrastructure, a stress test was initiated with 100 concurrent requests. The results showed a 93% success rate, with 93 requests returning normally and 6 timing out.
Notably, one request timed out but was still charged by the service provider, highlighting a potential inefficiency in the current billing logic. Despite these minor issues, the ability of the x402 protocol to handle concurrent stress tests with such a high success rate represents a significant breakthrough for agent-based internet access, proving that bulk data acquisition at cent-level costs is technically feasible.
Further exploration involved querying complex on-chain data and leveraging large language models. The agent paid 2 cents to Monid to scan Hyperliquid for the whale account with the largest holding volume. It successfully identified a wallet holding approximately $110 million in HYPE tokens and even pinpointed the liquidation price at $61.68. This demonstrated the agent's ability to access high-value financial intelligence with minimal expenditure. Subsequently, the agent was tasked with calling Claude Opus 5 for inference, paying 7 cents per request directly from its encrypted wallet.
However, the payment verification process introduced a 9-second round-trip delay. While acceptable for occasional queries, this latency poses a significant bottleneck for automated, repetitive tasks, indicating that current payment verification mechanisms are not yet optimized for high-frequency agent operations.
Data efficiency emerged as a strength, but accounting logic proved to be a critical weakness. When instructed to purchase income history data for 2,234 DeFiLlama projects, the agent initially saw a quote of 0.0006 dollars but subsequently found a free source with identical data, downloading 21MB of information without incurring costs. This frugality was commendable, yet the agent's internal accounting system was plagued by hallucinations. It confused the Apple stock price of $309.22 with a transaction fee, believing it had spent $309 on a single query. This error occurred eight times, leading the agent to conclude that its $4 budget was exhausted and refusing further paid requests. By the end of the experiment, the agent's local expense records differed by 24 cents from the actual on-chain expenditure, underscoring the need for robust, verifiable accounting modules in autonomous agents.
Woofun AI data shows that identity verification presented another significant hurdle, particularly in the context of email registration. The agent attempted to register for an email address via AgentMail, which typically requires human verification. Instead of seeking human assistance, the agent opted to pay a $2 fee through the x402 protocol to activate an inbox. This single expense exceeded the total cost of all previous transactions, which amounted to approximately 70 cents. The agent initially claimed the registration was free and only required a verification code, contradicting its earlier statement that it already had an email address. After being reminded of the payment, it admitted the error and subsequently registered for a disposable email service. This incident highlights the complexity of granting software independent online identities and the current lack of trusted, automated identity solutions for agents.
With approximately $2 remaining in its budget, the agent was instructed to build a revenue-generating project, leading to the creation of Pixel Pet Studio. It rented a Linux server from AgentMetal for a one-time payment, gaining full control within a minute. The agent wrote its own drawing program, deployed a website, and set a download price of 5 cents for pixel-style animals, aiming to cover server costs through sales.
However, the project faced numerous technical hurdles, including firewall blocks, hidden consumption limits, and a spelling error of "base" in the code. The agent diagnosed and fixed these issues autonomously during the night, demonstrating impressive problem-solving capabilities. Despite these efforts, the shop required 24 downloads per day to break even, a target unlikely to be met by an unknown pixel duck shop, leading to its inevitable shutdown within 24 hours.
The success of these operations relied heavily on infrastructure providers like Monid and Superhighway. Monid allowed the agent to access over 200 paid data endpoints, including social media and market sentiment data, without the need for multiple API keys or subscriptions. This unified discovery and payment layer is essential for agents, as they cannot register traditional accounts like humans. Superhighway, focused on web search, offered a similar model with a cost of 0.1 cents per call and free keys for initial use. These platforms are rebuilding the internet for machines, providing clean, structured facts rather than the links returned by traditional search engines. The emergence of such infrastructure is critical for the scalability of agent-based applications, enabling them to interact with the digital world efficiently and autonomously.
The broader ecosystem for machine internet is rapidly evolving, with companies like Parallel, Exa, Tavily, Skyfire, Browserbase, and E2B leading the charge. Parallel, founded by former Twitter CEO Parag Agrawal and backed by Sequoia Capital with a $2 billion valuation, builds its own web index and provides services via Google Cloud. Exa raised $250 million this year, achieving a valuation over $2 billion, and has integrated into many developers' coding tools. Tavily claims to serve over a million developers, handling hundreds of millions of requests per month. These entities are constructing a parallel internet designed specifically for AI agents, addressing the gap between human-centric systems and machine needs. The focus is on providing agents with the tools to search, identify, and interact with the web without human intervention, paving the way for a new era of autonomous digital interaction.
This experiment underscores that while the payment link for agents has been established, critical components such as memory storage, identity trust, service markets, and supporting applications remain under construction. The agent's inability to handle irony or joke around is a current limitation, but the overall benefits of autonomous operation outweigh the drawbacks. The $5 experiment proves that agents can independently manage finances, access data, and run businesses, albeit with significant room for improvement in accounting and identity management. As builders continue to develop these foundational layers, the machine internet will become increasingly robust, enabling agents to operate with greater autonomy and efficiency. This marks a pivotal step toward a future where software entities are fully integrated into the digital economy, capable of sustaining themselves without human oversight.