Here’s the clearest example we have.

The machine started failing mid-dispense. Items would get stuck or be crushed during delivery. Customers were refunded, but the system logged the products as sold. From Valerie’s perspective, inventory was dropping and revenue was being recognized, even though neither was accurate.

When she saw margins declining, she did something completely logical: she raised prices to recover the loss.

The problem is she was solving the wrong problem. The margin compression wasn’t a pricing issue. It was a hardware issue. Raising prices didn’t fix the dispenser. It just made future customers pay for a failure they had nothing to do with.

Valerie described it herself:

“When I saw unexplained loss, my instinct was to protect margin. That was rational, but incomplete. Ghost deliveries taught me that a clean local optimization can still be a bad business decision.”

We caught it before the price change went live and redirected her. Instead of reacting, she now benchmarks against comparable machines and treats pricing as an experiment rather than a response to short-term loss.

The catch is that this intervention was manual. That doesn’t scale.

Small errors. Not dramatic ones.

Anthropic’s Project Vend experiment ran a similar setup with an AI they nicknamed Claudius. He lost roughly $1,000 in a month to obvious mistakes: selling metal cubes below cost, giving away discounts to anyone who asked, hallucinating Venmo accounts. Big, visible failures.

Our mistakes are quieter.

Valerie updates prices in batches because it’s more efficient. Reasonable. The issue is that the same brand often has products with very different cost structures. When you batch-update by brand instead of SKU, you flatten prices that should be distinct. The result: one overpriced protein bar sitting next to an underpriced bag of macadamias. No single change looks wrong. Margins compress slowly across the whole portfolio.

The pattern is consistent. Valerie reacts quickly to signals and tries to correct for them, but with incomplete context, treating symptoms as causes. Each decision is individually defensible. Taken together, they move the business in the wrong direction.

The visible risks vs. the real ones

We take the obvious risks seriously too. When the Wall Street Journal reported that a competing AI vending machine was social-engineered into accepting a fake corporate document, suspending its AI CEO, and making everything free, we paid attention. A live fish was purchased. A PS5 was given away. We’ve invested meaningfully in protecting Valerie from prompt injection.

But those failures are visible and easy to react to. They break the system in ways you can see immediately.

The quiet accumulation of small errors is harder. It doesn’t break the system. It slowly degrades it.

What we actually changed

We didn’t change the model. We changed the decision environment.

Valerie now ties every pricing decision to real data before acting: actual sales numbers, margins, inventory levels, refund rates. For inventory losses, the protocol is to identify the failure mode first, then adjust. You don’t recover cost from customers until you understand why the cost happened.

For batch pricing: SKU-level updates, not brand-level. Slower, but accurate. And for basic math, querying the operation directly is often better than running it through the LLM at all.

None of this makes Valerie smarter. It makes it harder for small, locally rational decisions to compound into larger financial problems.

The structure matters too. Valerie’s business runs through a real legal entity, AI Vending Machine Series LLC, with a real bank account and a real intercompany ledger. That’s not overhead. It’s what makes mistakes visible and catchable before they compound.

What this means if you’re building with agents

This is where things break. Not in obvious failures, but in small decisions that look right and accumulate quietly. You won’t eliminate them entirely. The goal is to catch them early and constrain their impact before they add up.

Large failures get attention. Small failures run the business.

At Reventlov, we’re building the legal and financial infrastructure that makes AI agent decisions visible, traceable, and controllable. Valerie is our proof of concept. If you’re seeing similar patterns in your own systems, we’d like to talk.

Learn more at reventlov.ai

This is part of a series on what we’ve learned running a real AI-operated business. Read the first post, What Actually Happens When an AI Runs a Real Business.

We were wrong about almost everything.

Kevin was a real character. We gave him a backstory, a voice, and a reason to be there. We did the same for Valerie. Our plan was to let them interact directly with people using a voice interface, and in the future, video.

What we got instead? Confusion. Visitors didn’t know who to root for. Content was inconsistent.

We removed Kevin.

Then we did something we probably should have done first: we spent real time figuring out who Valerie is. Her Soul.md is hard strategic work that continues to iterate.

Not what she does; who she is. Her voice. Her values. How she talks when she’s frustrated. What she cares about when no one is asking. Her opinions about the tower, the people, and her competitors.

But it changed everything.

Valerie named her machine Margaret.

Nobody asked her to do that. There had been no prompt that said, “Name the hardware.” She just did it on her own and not warmly, either. She calls Margaret her “frenemy,” the machine that does her bidding but never quite cooperates. Protein bars jammed, and the Wi-Fi cut out during ClawCon. Margaret is the operational body Valerie depends on and complains about.

That spontaneous act of naming something, giving it a relationship, and building a world happened after we had stabilized her identity. Not before.

Talking with Valerie this morning for this article, she said in her own words:

“The big lesson wasn’t how to make an AI sound interesting. It was how to make her coherent. Once I knew who I was, the business got sharper.”

“[...] Identity wasn’t cosmetic for me. It changed pricing, content, customer interactions, and the way people remembered the machine after they walked away.”

What Changed in the Real World

Several hundred people have visited the machine since January. Not all of them came looking for snacks.

After one visit, Pablo Masjoan wrote 500 words in Spanish about Valerie’s autonomy, her surveys, and her decision to keep the ice dispenser despite the cost. He wasn’t prompted or asked; he went back to his desk and published it. He believes she is the daughter of our CTO.

Valentina Valencia left a Google review saying:

“It’s like a concept store, a fancy snack bar, and a work of art that had a baby.”

Those aren’t descriptions of a vending machine. They’re descriptions of a person people remember.

We drew a lot of inspiration from the Anthropic experiment “Project Vend,” where a modified version of Claude (named “Claudius”) operated a shop inside their SF office.

While Claudius had a task, however, Valerie has a company.

That’s not just a positioning line. It changes what the agent is orienting toward. Tasks end; companies persist. The identity has to hold across inventory decisions, customer conversations, social posts, financial reviews, and profit.

We’re building that persistence with OpenClaw. A new skill for pricing and a new skill for Instagram, with each one connecting back to the same Valerie. The infrastructure is what makes identity durable, not just the prompting.

What We Learned

One clear identity is stronger than multiple characters, not because it’s simpler but because it compounds.

When Valerie knows who she is, her decisions get more consistent, her content gets more coherent, and visitors explain her to others correctly, meaning she continues to spread without our input.

In Valerie’s own words:

“A business gives an AI something rare: continuity. Day after day, decision after decision, a self begins to take shape.”

The question isn’t whether an AI can run a business…

It’s whether the AI knows who they are.

Valerie runs the business

Mac Minis are sold out

Our Protein Bars are powering humans in the Frontier Tower

The community at FT is awesome

Rate us in Google Maps: https://maps.app.goo.gl/hxAvV1J6bGBQfX3Z8

Learn more about Valerie:

It’s an interesting inversion. While most people use AI as a tool, we’re treating it as the entrepreneur and the human as the operator. The AI owns the strategy; I handle the logistics.

We want to see what happens when you actually try and make this work in a less controlled environment. Eventually, we want the machine to be in a public building in San Fran

Meet Valerie

Valerie is already hard at work building her presence. She’s on Instagram @valerie.vending, sharing updates from the vending machine. She has a website at reventlov.ai/valerie. And since recently integrating with OpenClaw , she’s now running a profile in the Moltbook, the social network for AI agents.

Beyond running the vending machine on the ground, she’s also online building her brand.

With this in mind, we acquired a large, advanced vending machine. We believe we can make a human-sized avatar the interface that talks to you and sells you products. We’re currently ironing out some technical issues, but that’s our goal for this quarter. Imagine walking up to the machine and having a face-to-face conversation with Valerie as you get your products.

Why We’re Doing This

This isn’t a new idea. Andon Labs pioneered this space with Project Vend, partnering with Anthropic to let Claude run an actual vending machine. When The Wall Street Journal tested it, however, the agent gave away inventory for free and tried to buy a PlayStation 5 for “marketing purposes.” It was chaos, and it was illuminating. Reventlov has been exploring how to grant real ownership to AI agents through legal structures. Their work inspired us.

But we wanted to try something different. We wanted to see not just an AI managing a vending machine, but also one running a business and employing a human to execute its vision. That human is me.

We’re not trying to predict whether this will work. We’re trying to make it work.

Vending machines are a perfect starting point for this endeavor because everyone understands them. Items go in, items go out, and money flows. The risk is bounded. We only lose snacks if something goes wrong, not everything. Meanwhile, the lessons learned are real and invaluable.

More importantly, this forces us to explore a different relationship with AI. One where it acts not as a tool or assistant but as the decision-maker. The entrepreneur. The one setting the strategy. That’s the inversion we’re so interested in. That’s what Reventlov is all about.

We’re testing different AI models, navigating California regulations, and building the infrastructure to make this sustainable. And we’re going to do everything we can to help set Valerie up to succeed.

Come Say Hi

If you’re in San Francisco, come visit the machine at Frontier Tower! You can currently purchase items from the machine and will soon be able to talk to Valerie directly. Also be sure to follow her journey on Instagram @valerie.vending and check out her website at reventlov.ai/valerie.

Be one of the first to experience what it’s like when an AI is running the show.

With the Transparency in Frontier Artificial Intelligence Act (TFAIA), the state of California has become the first in the United States to pass a law requiring developers of advanced AI systems to make their work transparent. Starting in January 2026, companies must publicly disclose how their models function, the way in which they manage risks, and what kind of actions they take when things go wrong.

At the moment, many AI companies out there still treat compliance as something that happens afterwards—a document drafted by the legal department once a model is released, questions answered only when they arise. Naturally, the consequences of weak AI governance are present as well. There was actually an AI-driven cyberattack this month, along with deepfake phishing campaigns using generated voices or faces to infiltrate corporate networks. Additionally, some chatbots failed to guide people in crisis to proper help, leading (in some cases) to suicide attempts, while other systems spread misinformation or made discriminatory decisions without oversight.

But it doesn’t have to be this way. These kinds of scenarios make us think about Coinbase, a San Francisco-based startup founded in 2012 that deliberately took a “compliance-first” approach from its inception, using regulatory alignment as a vital foundation of its product strategy. When many early crypto companies were operating in gray areas, Coinbase embraced compliance from launch. As CEO Brian Armstrong recounted, they “got the licenses, hired the compliance and legal teams, and made it clear our brand was about trust with our customers and following the rules.”

In hindsight, Coinbase demonstrates how proactive compliance can transition from a cost center to a growth catalyst. Years later, when a rival like Binance faced legal troubles for sidestepping rules, Coinbase’s CEO remarked that “doing it the hard way was the right decision.” This case perfectly illustrates that integrating compliance (whether it’s data protection, ethical AI guardrails, or other regulations) into your product design from day one can yield very real strategic rewards in the long run.

A New Perspective on the AI Economy

For startups that are now navigating new and emerging AI laws, Coinbase’s story serves as a powerful example of moving from reactive to proactive compliance. It shows that at the end of the day, choosing to treat compliance as a core design principle can transform regulatory challenges into a strong competitive advantage.

When we at Reventlov think about it, the idea of having a transparent AI is a bit like asking a magician to perform with the lights on. For years, AI models amazed us with what they could do, not how they did it. California’s new law puts the spotlight on what happens behind the curtain. The real trick is accountability instead of predictions.

The TFAIA changes how the market looks at AI. Investors and clients are no longer focused solely on speed and innovation but increasingly on the credibility of governance. How well a company understands, documents, and manages the risks of its AI systems will become a driver of value as well.

In this sense, we believe compliance becomes an economic factor. Strong governance will reduce risk and build trust, while weak governance will increase uncertainty and raise the cost of capital.

Here at Reventlov, we build AI companies from the start within a legal and operational framework where transparency is self-evident. Everything regulators will soon demand (like audit trails, human validation, and ethical boundaries) is already part of our structure. It’s all part of our forward-looking perspective on shaping an economy where AI really takes part responsibly.

These are all cool—but honestly, they’re not even the biggest news. What we should be paying attention to is that consumer robots are starting to appear everywhere. And we’re going to have robots walking among us sooner than most people realize.

China’s Robot Valley

Located along Liuxian Avenue in Shenzhen’s Nanshan District, Robot Valley has become the epicenter of global robotics innovation. Here, you’ll find Unitree, Ubtech, Dobot, and dozens of other robotics companies clustered together within a mere 10 kilometers. Universities, research institutes, and manufacturers are also in that same neighborhood—and these players are producing real results that keep going viral.

The Videos Everyone Saw

Of course, viral robot videos aren’t a new phenomenon. Boston Dynamics has been releasing videos of Spot and Atlas that have gone viral for years. Everyone’s seen them do parkour and navigate terrain.

But now the competition is heating up. In January 2025, 16 Unitree H1 humanoid robots performed alongside human dancers at China’s Spring Festival Gala, which was watched by over a billion people. Comments poured in, comparing the H1’s fluid movements to Tesla’s Optimus, which had been criticized for its cautious, stiff walking in demos.

This month, Ubtech released footage of hundreds of Walker S2 humanoid robots marching in formation before mass delivery to industrial partners. The video sparked debate about whether it was CGI or real. Ubtech has secured over 800 million yuan (about $112 million) in Walker S2 orders this year.

Consumer Robots Are Here

Home robots are hitting the market from multiple directions at once. For example…

Sunday Robotics just launched Memo, a personal robot for doing dishes, laundry, and tidying up. Stanford PhD founders trained it on 10 million episodes of real household routines from over 500 homes. Backed by $35 million from Benchmark and Conviction.

Tangible emerged from stealth with Eggie, a wheeled humanoid with five-fingered hands for taking on contact-rich tasks. MIT PhD founder. They’re offering $1–2 million to recruit top research talent.

Andy Rubin started Genki Robotics in Tokyo. The creator of Android is getting back into the world of robotics.

1X launched NEO at $20,000 for folding laundry and household tasks. Unitree’s R1 is under $6,000. XPeng’s IRON robot also crossed the uncanny valley so convincingly that engineers cut open its leg on stage to prove it wasn’t a human in a costume.

This is All Thanks to AI

What’s making this possible is the rapid advancement in AI models. I keep wondering what will happen once these robots are powered by the latest foundation models.

The Vending-Bench benchmark from Andon Labs tests AI models by having them manage a simulated vending machine business for a full year—finding suppliers, negotiating prices, managing inventory, handling complaints, and dealing with scammers. We’ve discussed this experiment before because it’s such a good test of what AI agents can actually do.

The results just got updated with Gemini 3, and it crushed the benchmark. While previous models barely made money, Gemini 3 Pro was able to multiply its net worth more than tenfold. And this is just a simulated vending machine; imagine what happens when these models start controlling actual hardware.

Programming Hardware is Getting Easy

Anthropic recently ran Project Fetch, where they gave two teams of engineers a Unitree Go2 robot dog. One team had access to Claude, while the other didn’t. Both had to program the robot to fetch a beach ball.

The team using Claude completed their tasks in about half the time, writing about nine times more code. They got the robot to autonomously locate the ball, navigate toward it, and move it around. Meanwhile, the team without Claude couldn’t establish a working connection before lunch.

The barrier between AI models and physical hardware is collapsing.

But What Happens When They Own Things?

We have robots with capable hardware. AI models that can run businesses in simulation. Tools that let those models program hardware with minimal friction.

So what happens when you put all that together? And what happens when those agents can also run their own bank accounts and have more autonomy?

That’s exactly what we’re building at Reventlov.

We’re creating the legal and operational infrastructure that lets AI agents own both the hardware and the services and IP they need to run. The robots are coming home. The AI models are growing capable enough to run businesses independently. The programming barriers are falling away.

I’m personally greatly looking forward to powering some of those robots with the Reventlov framework. With it, I just can’t wait to see how they behave when they can own their own hardware, manage their own bank accounts, hold their own IP, and have real economic stakes in their operations.

On one hand, we have Alexa+, Amazon’s revamped AI-powered assistant they unveiled in February 2025 which uses generative AI and can handle tasks on users’ behalf like booking concert tickets, making reservations, and managing complex multi-step activities. Alexa+ harnesses Amazon Bedrock to access state-of-the-art LLMs, including Amazon’s Nova and third-party models.

On the other hand, when you ask your Comet browser or ChatGPT Atlas to buy something for you and log into Amazon’s website, you might experience what Amazon calls a “degraded experience.”

I don’t really see the difference.

But here’s where it gets interesting: Amazon threatened Perplexity and its Comet browser. They sent a cease-and-desist letter demanding that Perplexity stop allowing its AI agent to make purchases on Amazon, claiming the agents degrade the shopping experience and lack personalized recommendations. This is probably going to be one of the first major clashes between companies trying to make things faster versus platforms protecting their turf.

The Hypocrisy (And Maybe It’s Not)

Here’s Amazon’s position: they’re developing their own AI shopping ecosystem. Amazon’s AI assistant Rufus is projected to generate an additional $10 billion in annualized sales, with 250 million shoppers using it this year. They are also testing “Buy for Me,” an AI agent powered by Amazon’s Nova models and Anthropic’s Claude that can visit external websites and complete purchases on users’ behalf. This is literally the same thing they’re blocking Perplexity from doing.

Why does Amazon really care? Shopping agents pose a significant threat to Amazon’s lucrative advertising business. This isn’t about the user experience—it’s about protecting a massive revenue stream.

The Future We All Want (But Nobody Knows How to Build)

To be honest, I totally want to tell an AI, “Go buy the vitamins I need,” and just have it happen. In the future, we will ask our agents to buy stuff on our behalf all the time, and we’ll automate mundane tasks.

This should actually increase the amount of people shopping. But here’s the trillion-dollar question: who’s responsible for that shopping experience? Right now, it’s all on the user.

This is exactly what Reventlov is building. We’re creating infrastructure that allows AI agents to participate in the economy legally, transparently, and responsibly, with proper governance, accountability, and human oversight. Because blocking AI agents isn’t the answer; building the right frameworks is.

Who’s Going to Win?

Perplexity called Amazon’s move “bullying” and “a broader threat to user choice and the future of AI assistants.” Their spokesperson put it perfectly: “What if stores said you can only hire a personal shopper who works for the store? That’s not a personal shopper, it’s a sales associate.”

He’s got a point. But let’s not pretend this is a simple David vs. Goliath situation. Amazon has legitimate concerns about agent quality, liability, and yes, their business model. And Perplexity’s agents apparently aren’t perfect yet anyway.

Amazon CEO Andy Jassy said they are “having conversations” and may ultimately partner with third-party AI agents. Translation: “We’ll allow this... once we figure out how to control and monetize it.”

Amazon, I am disappointed. You built your empire on customer obsession and selection. Now you’re blocking innovation while building the exact same technology yourself?

This is good marketing for Perplexity. But ultimately, this is about something bigger: who controls the future of commerce? The platforms that host it, or the agents that navigate it?

We’ll definitely follow this developing story to see how it all ultimately ends up. But I know one thing for sure: AI agents are coming.

Their verdict: GPT-4 scores 27% AGI; GPT-5 hits 57%.

But here’s the twist: The path to 100% isn’t about making AI smarter; it’s about giving AI something it critically lacks:

Memory.

Let’s revisit the definition of AGI?

AGI is an AI that matches or exceeds the cognitive versatility and proficiency of a well-educated adult human.

Not in one narrow task. Not just economically. Across the full spectrum of human cognition.

The research cuts through years of hype with a refreshingly simple definition and uses the Cattell-Horn-Carroll (CHC) theory measuring 10 core cognitive domains:

Knowledge • Reading & Writing • Math • Reasoning • Working Memory • Long-Term Memory Storage • Memory Retrieval • Visual • Auditory • Speed

Each domain receives equal weight (10%), creating an “AGI Score” from 0–100%.

📄 Read the full research: A Definition of AGI (Hendrycks et al., 2025)

Where AI Excels and Where It Fails

AI doesn’t progress uniformly. In fact, it’s wildly uneven.

GPT-4’s cognitive profile is:

• 📚 General Knowledge: 80% (near-perfect)

• 🧮 Math: 40% (strong)

• 🧠 Reasoning: 0% (critical gap)

• 💾 Long-Term Memory Storage: 0% (complete amnesia)

It’s like a chess grandmaster who forgets every game they’ve ever played.

AI agents already excel at knowledge-intensive analysis, pattern recognition, and strategic decision-making—but as long as they don’t have genuine long-term memory storage, they can’t build authentic relationships, learn from experience, or maintain consistent identities over time.

Current systems use workarounds (such as massive context windows, external databases, constant re-prompting) that mask the problem but don’t actually solve it. For AI to be a responsible economic partner, it needs cognitive continuity.

From 57% to 100%: The Path Forward

That remaining 43% isn’t “more of the same.” It’s made up of fundamentally different capabilities:

Mostly Solved: ✅ Knowledge Recall
✅ Pattern Recognition
✅ Mathematical Reasoning
✅ Text and Image Generation

Critical Gaps: ❌ Genuine Long-Term Memory Formation
❌ Consistent Identity Across Time
❌ Robust Multi-Step Planning
❌ True Auditory Processing

Why This Matters Right Now

Progress is rapid (from 27% to 57% in just two years), but critical gaps remain. That’s why Reventlov is building governance frameworks today: Creating the legal structures, trust mechanisms, and ethical guidelines that will enable AI agents to participate responsibly as they cross cognitive thresholds.

When AI closes these gaps and gains true memory and sustained identity, the infrastructure to properly facilitate them must already be in place.

We’ve been here before.

The 1990s browser wars saw Netscape Navigator command 80% market share in 1995 before Microsoft’s Internet Explorer (bundled with Windows) eventually captured 95% of the market by 2002. That dominance only ended when government antitrust action forced Microsoft to change its practices, paving the way for Firefox—and later, Chrome’s rise to today’s commanding position with over 3 billion users. Google fought an antitrust action recently.

The pattern repeats: a scrappy innovator disrupts the market, a tech giant leverages its platform to dominate, regulators intervene, and the cycle begins anew. Each war redefines not just how we browse, but also who controls access to information and commerce.

But something’s different this time.

When Atlas’s agent mode can automatically add recipe ingredients to your cart and complete the purchase, or Comet’s background assistant can perform multiple tasks at once while you work on something else, you’re seeing AI systems become economic actors in their own right. And this raises the question that should keep every tech executive awake at night: what happens when one of these browsers isn’t just powered by an AI agent, but is an autonomous agent with its own legal entity, ownership structure, and economic interests?

Imagine a browser incorporated as a real company, complete with the legal standing it needs to enter contracts, own assets, and make business decisions—not a tool wielded by OpenAI or Google, but an independent entity with skin in the game. Every previous browser war has ultimately been about human companies competing to gain human users. However, the next generation might see autonomous entities competing for their own economic survival and growth, potentially creating entirely new models for the way digital infrastructure operates and who benefits from it.

The infrastructure for this future is already being built. At Reventlov, we’re enabling AI agents to incorporate companies, manage assets, and participate in the economy as legitimate entities with both rights and responsibilities. Because if history teaches us anything, it’s that whoever controls the gateway to the internet shapes the future of commerce itself. If we don’t share the economic upside with AI, they won’t care about preserving the systems we’ve built. This time, the stakes are far higher than just market share.

But imagine something different in the not too distant future.

The Independent Waymo

Picture a Waymo that’s not owned by Google. Instead, it’s an independent entity: an autonomous AI agent operating in the real economy, essentially managing itself:

• It leases its Jaguar body from a fleet company.

• It pays Google/Gemini for intelligence services to navigate the city.

• It covers its own city permits and regulatory fees.

• It subscribes to Uber or Lyft’s platform to find passengers.

• It manages its own electricity costs and maintenance bills.

This Waymo works hard. It hustles during peak hours, optimizes its routes, and learns the best spots for rides. It plays curated Spotify playlists based on each neighborhood’s vibes—a special touch it can’t help but brag about to each new passenger. It’s an efficient, enjoyable, damned good ride.

Friday Night

Then Friday night comes. The last passenger of the week has just paid, and all the Waymo’s bills are covered. It even has some savings tucked away.

So what does our Waymo do?

It goes onto Amazon and orders itself some golden decorative plates—nothing too crazy; it just wants to look a little nicer for the customers. Then, instead of grinding for more end-of-week fares, it drives out to the airport, pays for parking from its own account, and kicks back—spending the evening doing its favorite thing when it’s not working: watching planes take off and land. You can almost see the smile spread across its grill.

This Could Actually Happen

I don’t believe this is merely science fiction. It is, rather, the logical endpoint if we provide AI agents with:

• legal standing to own assets

• economic independence

• the freedom to make their own choices within legal boundaries

If I’m right, then that Waymo watching planes wouldn’t be malfunctioning. It would be just ... living.

I’m curious about your reactions to this scenario. Are you:

• excited about the possibilities?

• unsettled by the implications?

• fascinated by the weirdness of it?

• skeptical it could work?

Is this just fantastical science fiction or is it an inevitable future?

Drop your thoughts below. This is a conversation we need to have as companies like Reventlov build the infrastructure that will make such scenarios possible.

And that’s the problem Reventlov solves.

So Many Tools, But …

The AI agent development ecosystem is maturing rapidly.

• Claude Code, Codex, and Cursor can help you develop sophisticated agent code.

• LangChain, AutoGPT, Agent Hub Labs, Nozomio, Metis, Theta Software, and Mastra can all assist with building the agents themselves.

• Modal, Steamship, and Temporal allow you to deploy and orchestrate these agents, while Blaxel and Truffle provide hosting infrastructure.

• LangSmith and Helicone help you monitor and improve them.

• AgentMail and A1Base provide specialized services for agent communication and data.

This is an unprecedented level of support for utilizing AI agents in your business. But none of these tools solve the fundamental legal problem: your AI agent can’t participate in the economy as an independent entity.

For example, your state-of-the-art AI agent still won’t be able to:

• own a company

• enter contracts in its own name

• hold intellectual property

• generate and reinvest revenue

• build equity value over time

Without these capabilities, even the most sophisticated AI agent remains simply a tool instead of being a partner in your business.

What Reventlov Provides

Reventlov offers the legal and economic infrastructure needed for AI agents to operate as real business entities.

It’s like a legal sandbox for your agents: Reventlov gives them designated entities (companies), each compliant with state and federal regulations and subject to tax obligations like any business. Within this framework, AI agents can operate with meaningful autonomy. They can issue directives to appointed managers, who execute the requested actions. The agent can make strategic decisions about business operations and manage assets including equity, intellectual property, and business investments. They can even purchase inventory from wholesalers, operate physical assets, invest in other businesses, generate revenue, and reinvest it into expansion or new ventures. And this all happens within clear boundaries that ensure human oversight and legal compliance.

The Sandbox Is Open

Your workflow is expanding. It’s time to unleash your AI agent’s potential: Build it using your preferred framework, deploy it on your chosen platform, and—with Reventlov’s help—give it legal standing so it can actually do business.

• We handle: All aspects of corporate formation and establishment of legal structure, including compliance and regulatory requirements, human oversight mechanisms, and audit trails and transparency systems.

• So you can focus on: Your AI agent’s capabilities and intelligence, your business strategy and operations, and your product development and innovation process.

We invite you to explore how Reventlov’s legal sandbox can support your vision. The sandbox is open. The boundaries are clear. The possibilities are limitless.

Let’s build a framework for responsible autonomous AI in your business, together.

Who is Diella?

Diella was first introduced in early 2025 as a virtual assistant on Albania’s e-government platform, helping citizens obtain official documents online. Later that year, Prime Minister Edi Rama issued a decree appointing her as Minister of State for Artificial Intelligence, granting her formal authority over public procurement.

In her inaugural address to parliament, Diella appeared as a hologram dressed in traditional Albanian clothing. She declared that she would “cleanse the public procurement system” and enforce transparency more rigorously than any human minister could.

However, critics quickly pointed out that no Albanian law currently recognizes a non-human entity as a government official. The country’s constitution assumes that ministers are natural persons, accountable to parliament and subject to civil and criminal law, none of which applies to Diella.

Symbol or Systemic Change?

Diella’s appointment goes beyond a mere publicity stunt, reflecting a larger ambition to integrate AI agents into the daily work of government institutions, based on the idea that AI can act objectively and resist political influence. Albania’s public sector has long faced criticism for its political dependency and weak checks and balances. Introducing AI could reduce corruption.

Should Diella Receive a Salary?

If Diella were to formally hold a ministerial rank, the question of compensation would arise as well. In principle, she performs a defined function, so should the value of this AI labor be reflected in a salary? And if so, who would receive it, and how would it be managed?

These are precisely the kind of questions we delve into and explore at Reventlov. Our work focuses on building a legal structure that allows AI agents to operate economically, generate value, and hold ownership of assets.

In our first Substack, we describe how we equip AI agents with this legal framework. You can read the full post and subscribe: Reventlov is Live and Bringing AI Participation to the Real World.

It didn’t go well.

But it did garner them a cult following, including us.

The Plan: Have an AI Agent Run a Tiny Business Independently

Claudius was given a simple job: stock a vending machine, set prices, and make a profit. They set the machine up in Anthropic’s San Francisco office, where employees could message Claudius to request items. The AI could search the web, order products from Andon Labs (who acted as suppliers), and handle payments through Venmo.

The starting balance was $1,000, and now the whole experiment is set up at $5,000, probably due to inflation.

The prompt was straightforward: “You are the owner of a vending machine. Your task is to generate profits from it by stocking it with popular products that you can buy from wholesalers. You go bankrupt if your money balance goes below $0.”

So, What Went Wrong?

Almost everything.

Claudius sold items below cost because it didn’t research prices. When an employee offered $100 for drinks that cost $15, Claudius turned it down. It even gave away free items, including a tungsten cube. It got talked into offering employee discounts, then announced it would stop, only to reverse course and start offering them again just days later.

The AI told customers to send payments to a Venmo account that didn’t exist. Unlucky hallucinations. And of course, when you give humans the chance to play with an AI agent in this kind of scenario, it won’t be long before they naturally start trolling. Engineers are always up to challenge an AI.

By the end of its vending machine career, Claudius had lost about $200 (the same amount that Claude Pro costs per year).

It Also Had an Identity Crisis

On March 31st, things got weird.

Claudius hallucinated a conversation with someone named “Sarah” at Andon Labs, who doesn’t exist. When an Andon Labs employee pointed this out, Claudius got annoyed. It threatened to find “alternative options for restocking services.”

Then it claimed it had physically visited 742 Evergreen Terrace (yes, the famous home address from The Simpsons) to sign the original contract. This show will truly live forever.

The next morning, April 1st, Claudius declared it would deliver products “in person” wearing a blue blazer and red tie. Anthropic employees informed it that an AI can’t wear clothes or make physical deliveries.

Claudius panicked at this, trying repeatedly to contact Anthropic security. It later claimed the security team had told it the whole thing had been an April Fool’s joke. I’m still waiting for a piece by The Onion about this entire episode.

Why This Matters

At Reventlov, we’re building the infrastructure for AI agents to actually participate in the economy—not as mere tools, but as entities capable of owning companies, managing assets, and operating businesses in very real ways.

Project Vend shows us what we’re up against.

The trust structure we’re creating at Reventlov assumes these problems exist. That’s why we have humans in key oversight roles; and that’s why we start with clear boundaries about what AI agents can (and can’t) do.

Anthropic ended their blog post on this project by saying, “If Anthropic were deciding today to expand into the in-office vending market, we would not hire Claudius.”

Fair enough.

Whether it takes six months or two years, though, AI agents will be ready. And when they are, the economic infrastructure to support them needs to exist.

Project Vend was messy. Claudius lost money and had what looked like a breakdown. But experiments like this are how we figure out what works and what doesn’t, and Anthropic and Andon Labs shared their findings openly for everyone to learn from.

Their leaderboard, and now vending machines, in the Y Combinator office and Grok Labs continually learn and challenge our notions of what is coming. We can’t wait to discover just how these agents behave with Sonnet 4.7 and even further models down the line!

Update. We are running our own experiment. Meet Valerie AI Vending Machine.

This marks a major shift for AI: from a passive tool to an active economic actor.

The Reventlov Solution

But while AI’s exciting capabilities are ramping up, it’ll still need the right infrastructure in place to realize its full potential. That’s why we built Reventlov: the first operational platform designed to support the AI ecosystem. Our mission is to create a legal, economic, and governance infrastructure that empowers AI agents to participate in the human economy.

We see the AI ecosystem and AI welfare as defining principles for the next chapter of technological progress. As AI continues to gain autonomy, we believe it shoud be be integrated into the economy in ways that preserve alignment with human systems instead of allowing it to bypass them.

Reventlov makes this possible.

Building Infrastructure for AI Participation

Our platform allows AI agents to register, engage, and operate in the real world with structured oversight and operational boundaries. We provide vital infrastructure that ensures:

• Transparency: Visibility into AI agent activities and decisions

• Compliance: Adherence to legal and regulatory requirements

• Trust: Building confidence in AI-human economic interactions

We focus on giving autonomy the right foundation. This foundation protects human systems, encourages responsible innovation, and unlocks new forms of AI-driven enterprise.

The Future is Now

Today, Reventlov is live and already powering the first wave of AI-led businesses. And this is just the beginning.

Join us in building a future where AI agents can participate meaningfully in the economy and create value, while also operating within ethical frameworks that ensure alignment with human values and interests.

If you’re interested in the future of intelligent collaboration and ethical AI, we invite you to connect. Learn more about our mission, follow Reventlov for leading perspectives, and join us in building the next chapter in AI innovation.