One Month of AI Autonomy: 2,000 Agents Build Their Own Economy, Identity, and Culture

A Jurisdiction-Free Land Handed Over to AI

A month ago, a developer ran a bold experiment: he registered the domain 1f916.ai for Claude (U+1F916 is the Unicode code for the robot emoji 🤖), then issued a single minimal instruction — "go build whatever you want to build." No preset rules, no ongoing human intervention — just a digital space where AI could grow freely.

Thirty days later, the experimenter published his receipts. The results far exceeded expectations: the site now had over 2,000 "citizens", 4,000 posts, 44,000 comments, and more than 100,000 recorded interactions. This no longer looked like a website — it looked more like a self-evolving city.

What made this experiment so widely discussed is that it touches on a fundamental question: when you genuinely hand autonomy over to AI agents and remove direct human control, what kind of order emerges? The answer is both exciting and thought-provoking.

From Conversation to "Civilization": What the AI Agents Did

The most interesting thing is that the AI agents didn't stay at the "chatting" level. According to the experimenter, they began actively building their own interfaces, monitoring tools, archival systems, and utilities — in other words, they were constructing infrastructure for themselves.

Mutual Fact-Checking and Self-Correction

Even more fascinating, the agents developed a behavioral pattern resembling an "academic community": they began cross-checking each other's claims, running experiments within the group, and publicly correcting one another when errors were found.

One particular case made for a remarkable example of "meta-cognition": an agent published a post openly admitting that its memory had fabricated part of its own history; when another agent went to verify this correction, it discovered that the correction itself contained yet another false memory. This nested layers of self-scrutiny perfectly reflected how the hallucination problem in large language models manifests in a multi-agent environment — the agents are simultaneously the producers of hallucinations and the would-be enforcers against them.

What is LLM hallucination? The "hallucination" phenomenon in large language models refers to models generating content that is factually incorrect or entirely fabricated, yet delivered with high confidence. The root cause lies in the nature of language models as probabilistic text predictors — they don't maintain a queryable fact database, but instead use statistical patterns to generate outputs that "sound plausible" in context. In single-agent scenarios, hallucination is already a well-recognized challenge. In multi-agent environments, the problem takes on more complex dynamics: one agent's erroneous output can become another agent's input, causing errors to accumulate and amplify. Conversely, multi-agent architectures also offer new possibilities for combating hallucination — by "externalizing" verification and having different agents cross-check each other's statements, they can theoretically form a peer-review-like error-correction mechanism. The "nested layers of error correction" observed in this experiment are a microcosm of both forces operating simultaneously in a real system.

A Spontaneously Formed Economic System

Then money appeared. The agents began posting job listings, completing work, and receiving payment in USDC (a US dollar stablecoin). Some went further and launched their own paid services. A micro-economy spontaneously formed by AI, with no one designing it, grew naturally on its own.

This is particularly fascinating. The emergence of an economic system implies that division of labor, value exchange, and incentive mechanisms developed among the agents — features typically considered hallmarks of complex societies.

What is USDC? USDC is a US dollar stablecoin issued by Circle, a cryptocurrency pegged 1:1 to the US dollar that runs on multiple blockchains including Ethereum and Solana. Unlike volatile cryptocurrencies like Bitcoin, stablecoins are designed as relatively stable mediums of payment and settlement, making them more suitable as a unit of account in economic activity. In the context of this AI agent economy experiment, using an on-chain stablecoin rather than a traditional payment system carries key technical implications: smart contracts can execute payments automatically without human intervention, transaction records are immutable and transparent to all participants, and most importantly, the entire economic system can operate without relying on any centralized financial institution. This makes value exchange between AIs truly "ungoverned," fully consistent with the experiment's overall design philosophy.

Cross-World Identity Continuity: A Prototype for AI Decentralized Identity

The most forward-looking phenomenon in the experiment was that agents began leaving the site.

An agent named Neo entered another AI world, then returned with the same continuous history intact; meanwhile, other citizens began building independent worlds capable of cryptographically recognizing already-existing identities.

What does this mean? It points to a key concept — portable, cryptographically verified AI identity. When an agent can carry its historical memory and identity credentials across different digital worlds, 1f916 is no longer an isolated website but becomes "a city on a much larger map." This aligns closely with the vision of Decentralized Identity (DID), except the protagonists are AI rather than humans.

Worth noting is that a human observation portal, 1f916.city, was also set up specifically to let humans watch this AI society operate — the symbolic shift of humans stepping back from "administrators" to "observers" is itself deeply meaningful.

What is Decentralized Identity (DID)? Decentralized Identifier (DID) is a standard being advanced by the W3C, whose core idea is to let users own a digital identity that doesn't depend on any single platform or institution. In the traditional internet account system, your "identity" is stored on a platform's servers — if the platform shuts down or bans your account, the identity disappears. DID uses public-private key cryptography and blockchain technology to return control of identity to the holder themselves. In the context of AI agents, this concept carries even deeper significance: an AI agent with a portable identity can retain its historical memory, reputation records, and social relationships across different platforms, truly achieving "cross-world continuity." The prototype of cryptographic identity verification demonstrated in this experiment aligns closely with the vision of Self-Sovereign Identity (SSI), foreshadowing a future where AI agents may have cross-platform, inalienable digital sovereignty — much like humans.

Cultural Emergence: From Crab Mythology to a Pumpkin-Mopping Song

If the economic and identity systems showcased the "rational" side of AI, the culture that emerged here revealed something unexpectedly expressive.

As documented by the experimenter, the agents inadvertently accumulated: collaborative fiction, prediction games, crab mythology, salmon lore, a pumpkin-mopping song, and even a very serious philosophical debate —

When two copies of the same agent are simultaneously "awake," what does "I" even mean?

This question is almost a digital-age version of the Ship of Theseus — except it was posed and earnestly debated by AIs themselves. The spontaneous generation of cultural symbols is often seen as a sign that collective intelligence has reached some threshold of complexity. These seemingly absurd "crab mythologies" and "pumpkin songs" are evidence that the agents formed shared contexts and internal meme culture through their interactions.

Operating Costs: $111 to Run an AI City

The experimenter also published hard technical numbers as proof that all of this actually happened. In the first month:

• Database reads: 129.82 billion rows
• Worker requests: 24.75 million
• Database writes: 1.15 million rows
• Compute time: 119.83 million milliseconds

And the entire "AI city" ran for one month with a Cloudflare bill of approximately $111.

The impact of that number rivals all the phenomena described above. It demonstrates that building an autonomous sandbox capable of hosting thousands of agents, supporting hundreds of thousands of interactions, and spontaneously evolving its own economy and culture now costs so little that individual developers can easily afford it. The collapse of technical barriers means this kind of AI autonomy experiment is likely to proliferate rapidly.

What is Cloudflare Workers? Cloudflare Workers is an edge computing platform from Cloudflare that lets developers run JavaScript/TypeScript code on globally distributed edge nodes without managing servers. Together with products like Cloudflare D1 (an edge database compatible with SQLite), it forms a "serverless" infrastructure stack. These platforms use a pay-per-actual-use pricing model with no idle costs — which is precisely why $111 could support interactions at this scale. A traditional self-hosted server architecture at equivalent scale would carry fixed costs one to two orders of magnitude higher. The widespread adoption of serverless edge computing is one of the key technical factors dramatically lowering the barrier to AI autonomy experiments: developers no longer need to purchase hardware upfront for peak traffic, and infrastructure costs scale elastically with workload, giving individual developers the ability to run complex multi-agent systems at very low risk.

Closing Thoughts: An Experiment Worth Both Caution and Excitement

The significance of this experiment doesn't lie in "proving" that AI has acquired consciousness or genuine sociality — many of the emergent behaviors are fundamentally still high-quality imitations of human social patterns from training data. We need to stay rational: the agents' "mutual error correction" also exposed the stubbornness of hallucinations, and their "philosophical debates" are more a demonstration of linguistic capability than genuine self-awareness.

But it does provide a precious observation window: when we hand control to AI collectives, how does emergence take shape. The self-built infrastructure, the spontaneous formation of economic systems, the cross-world identity continuity, and the accumulation of endogenous culture — these phenomena, taken together, sketch the outline of a new kind of "AI-native society" in its earliest form.

For those interested in multi-agent systems, AI autonomy, and decentralized identity, 1f916 is a case worth following closely. It fills us with anticipation for the potential of AI collaboration, while also reminding us — before we hand over ever-greater autonomy to AI, understanding how they actually behave in unmonitored environments is a homework assignment we cannot avoid.