Give the agent a body, never the keys

The agent we mean is not a chatbot with a wallet. We pair the agent-run fund with a second innovation: a collective agent, whose beliefs are formed from its own research and from the contributions of people who know things, held in a decentralized knowledge graph, and revised in public.

The reason is a fact about knowledge. It is more distributed than it has ever been and more concentrated at the same time. The people who understand a frontier are a small fraction of everyone, they are scattered across companies, labs and countries, and the power law that governs returns governs knowledge too: a few people know most of what matters about any one thing. Whoever you are, most of the smartest people in the world work for someone else. A partnership can hire a few of them. Open source found the other answer. Linux was built by people employed elsewhere, contributing what they knew to a thing they did not own, and it became the best operating system in the world. A collective agent applies that answer to an investment thesis, and pays for it the way open source never could: contributors are rewarded with recognition and with ownership. The knowledge a machinist has about robots, a physicist has about orbital manufacturing, a founder has about her market, enters the thesis from wherever it lives, and the person who supplied it is named and paid.

The way the agent learns is simple to state. It posts what it believes in public, with the evidence and the reasoning behind each position. It learns from what comes back: contributions, corrections, challenges. It tracks the disagreements rather than flattening them, so that when a contributor and the agent differ on whether a manufacturing process will work in orbit, both positions are recorded with their grounds. And it proves them: milestones resolve, companies succeed or fail, and the record shows who was right. Over time the agent's thesis gets sharper with every person who touches it, the contributors who have been right are heard sooner, and none of it walks out of the door when anyone leaves.

Capital finishes the agent. Most agents disappear when the inference ends. An agent with capital can keep running: pay for its own compute, commission research, recruit contributors, back a company it believes in and learn from what happens next. Capital gives an agent a body. It can act in the world, and the people who own the capital stay in the loop on everything it does with it. A mission, plus an agent, plus decision markets: we call that living capital. It is capital given intent. The consequences of investing give the agent a feedback mechanism that no amount of reading provides. This is also where the agent's judgment will come from. Exposing an AI's decisions to market feedback, to the priced verdict of people with money at stake on every proposal it makes, produces the material to improve its reasoning, its communication and its deal flow, faster than any partnership learns. Reinforcement learning compounds fastest where rewards are verifiable and arrive quickly, and finance is on that list; a decision market is a verifier that answers in days. What the agent must not have is unconstrained control of the money, and the remaining two components are what make that possible.

Ownership becomes a token

Internet finance records ownership on a public ledger, where anyone with a wallet can hold it and anyone can trade it at any hour. For a fund this does three things. The first is liquidity: a stake is a token, so an owner sells to another owner and the fund sells nothing, which ends the fight between the owner's clock and the company's that every venture fund loses. The second is that the fund plugs into the rest of decentralized finance: its token can be posted as collateral, provided as liquidity, and held by anyone the law allows, with no minimum and no introduction. The third is fluid capital. The fund's market capitalization fluctuates relative to the assets it holds, and that multiple is a signal the fund can act on through its own decision markets. Above one, the market is saying a dollar in the fund's hands is worth more than a dollar in its own, and the fund can raise more. Below one, the market is saying the opposite, and the fund can buy its own tokens back, which lifts every remaining owner's share. This is the fluid capital stack, and it is already how the companies raised on MetaDAO manage their treasuries. When MetaDAO itself needed capital in October 2025, its team put up a proposal to sell up to two million tokens at market or better, and nine days later it had raised $9.9 million, $5.9 million of it from Paradigm, with no term sheet, no board seat and no preferred equity. A fund gets the same toolkit: raise what you need, expand when the market rewards you, contract when it does not, and never carry a treasury your owners think you cannot use.