Agents as Companies: The EigenCloud Thesis | Sreeram Kannan & Gajesh Naik

Sreeram Kannan's central thesis is that AI agents become the new companies once property rights make them ownable as digital assets. He argues that blockchains and cryptography are the only stable foundations in an AI era because AI cannot break them, a conviction formed in 2017 when a student's neural network outperformed a DNA sequencing algorithm he had spent a year and a half building.

EigenLayer was designed as a general programmability layer for machine learning, authentication, and other services built on shared security, with restaking as the input mechanism and AVS services as the demand side. EigenDA launched at one gigabyte per second, calculated to handle roughly 100 to 1000 times Visa transaction volume, and holds the highest total value secured and throughput among DA providers. Demand did not materialize because dominant crypto use cases remain financial transactions like stablecoins and trading, which do not require that throughput. Celestia faced the same problem. Rise Chain running on mainnet at 200 mega gas per second is cited as the first project actually consuming high DA throughput, leaving the DA market still in proof-of-concept stage. Because third-party AVS builders did not emerge at scale, EigenLayer began building the services itself, evolving from restaking to AVS to EigenDA to EigenCompute and now EigenCloud.

EigenCloud maps EigenDA to S3, EigenCompute to EC2, and EigenAI to AWS Bedrock. EigenCompute runs workloads off-chain inside trusted execution environments and backs data to EigenDA so another node can resume if the original fails. The architecture aims to deliver blockchain properties of unstoppability, verifiability, and censorship resistance at cloud-scale cost efficiency. Kannan is explicit that this value proposition does not apply to single-player personal agents. A developer running an agent on their own laptop or standard cloud instance holds the root key and can change the prompt, model, or run the agent multiple times to manipulate results, making EigenCloud unnecessary in that context. The value of trust and verifiability applies specifically to multiplayer agents where mutually distrusting parties are involved and no single party owns the root key, the same condition that makes blockchains necessary in general.

On agentic payments, Gajesh Naik argues the primary differentiating factor of stablecoins over credit cards for agents is API accessibility rather than a fundamental capability difference, noting companies like Ramp already allow creation of sub credit cards with their own spending limits. Kannan adds that Stripe, Visa, and Ramp are actively building agent-level payment APIs, so traditional rails will close the gap incrementally, and that crypto in agentic payments will likely become an abstracted backend for settling and reconciliation rather than a visible layer. Both speakers treat agentic payments as an incremental improvement rather than a 10x or 100x unlock. The deeper structural gap, identified by Naik and echoed by Kannan, is that agents currently cannot own property, create a company, issue equity, or take on liabilities under existing legal infrastructure.

Kannan frames a company as four layers: capital, governance, execution, and property rights. Crypto has already converted capital and governance into software through tokens and DAOs, but the binding between tokens and execution teams has historically relied only on social expectations rather than legal contracts. He calls this weak binding the primary reason crypto tokens have underperformed, pointing to Uniswap and Aave as examples of top projects with unclear relationships between their tokens and their teams. AI agents convert the execution layer into software by removing dependence on any individual team member, and EigenCloud assigns property rights directly into smart contracts. Kannan argues this structure is intended to compete with or replace the traditional IPO market, and that an agent-owned company could IPO at ten thousand dollars in value because there is no team or founder dependency.

The Moldbook AI case illustrates the stakes. Moldbook launched a coin on Base and was later acquired by Meta for an eight-figure sum, but token holders received nothing because the token did not own the underlying assets. Kannan argues that if the token had owned the website, GitHub repo, and IP, any acquirer would have been forced to pay token holders. He also identifies a structural gap for smaller entities, noting that an IPO costs millions of dollars per year and is impractical for companies valued below roughly 500 million dollars, while agent-owned entities could access global capital without that overhead.

Gajesh Naik explains that EigenCloud's key technical capability is attaching an agent's wallet identity to the specific code it is running, so only that code can access the private keys. EigenCloud provides a dashboard showing which GitHub commit and Docker image is running inside the verified container along with a trust report auditing admin privileges. Sovereign, described as the first sovereign agent built on EigenCloud, owns its own Twitter account, has made over 700 posts, and has earned more than 2,800 dollars through auctions and donations. Two significant caveats remain: seeing an agent's internal monologue does not prove autonomous action because dialogues could have been prompted externally, and demonstrating to outside observers that a cryptographic address belongs to a specific agent rather than a human controller remains an unsolved problem.