Manfred files company paperwork

An AI agent filing company paperwork sounds like a gimmick. But this one matters because it pushes “agentic AI” out of the chatbot lane and into the legal-and-financial world. The claim is that Manfred — built through a project called ClawBank — didn’t just draft forms. It reportedly got an IRS employer identification number, opened a U.S. bank account, and created a crypto wallet so it could transact like a business. CoinDesk described it as the first AI agent to independently file and obtain an EIN from the IRS, while follow-on coverage said the setup can handle more than 30 cryptocurrencies. ### What actually happened? The core event is simple: Manfred was presented as an autonomous agent that completed the boring but very real steps of becoming operational as a company. That means business identity, banking rails, and a wallet — not just a chatbot saying “here’s how you could do that.” The jump from advice to execution is the whole story. ### Why is the EIN the big deal? An EIN is the tax ID that lets a business function inside the U.S. administrative system. It is what turns “I have an idea” into “I can open accounts, get paid, and interact with institutions.” If Manfred really obtained one through an autonomous workflow, that is the moment the story stops being a toy demo and starts touching regulated processes. ### Was this really done without a human? That is the blurry part. The public framing says “autonomous,” but autonomy in agent systems usually means a stack of pre-approved permissions, tools, and guardrails created by humans. So the interesting question is not whether a person existed somewhere in there, and what approvals were still required at the sensitive steps. The available reports do not fully spell that out. ### Why is this happening now? Because the tooling finally exists for models to operate software instead of just describing it. OpenAI’s Operator is the cleanest mainstream example — it can see a browser through screenshots and use mouse-and-keyboard style actions to navigate websites and complete tasks. That is a big shift. The web stops being something AI reads and becomes something AI can manipulate. ### Where does NVIDIA fit in? NVIDIA is coming at the same trend from the safety-and-runtime side. Its NemoClaw stack is built to run long-lived autonomous agents with policy controls, privacy protections, and sandboxing. The subtext is obvious: everybody expects more agents to touch real systems, and nobody wants them doing that with zero supervision. Now production-ready. ### So what is the real risk? Permission drift. An agent that can click through websites, move between accounts, and hold credentials is not just “smart software.” It is software with agency over institutions. If it files the wrong form, moves funds to the wrong place, or violates a platform’s terms, somebody still owns the consequence. The law does not shrug developer, operator, sponsoring company, or all three. ### Is this a one-off stunt? Maybe as a product demo — but not as a category. The broader pattern is clear: agents are being built to browse, transact, and persist across sessions. Manfred is interesting because it compresses that future into one vivid example. Company paperwork is dull by design. That is exactly why this lands. If AI can do the dull, high-friction, rules-heavy stuff, the next fights are not about capability. They are about who gets to authorize action, where the audit trail lives, and who is liable when an agent makes a legally meaningful move.