Most AI agents are maintained by someone. A server has to be paid for, a database kept running, keys held somewhere. The moment that maintenance stops, the agent stops.
The Secret Network team built this proof-of-concept (PoC) to explore what a self-sustaining agent could look like by pairing confidential compute and persistent memory. What follows is a walkthrough of how it works, and how Secret Network’s SecretVM and Autonomys’ Auto Drive can give AI agents greater autonomy while maintaining a permanent, verifiable record of their activity.
Secret Network is the first mainnet blockchain with privacy-preserving smart contracts, launched in 2020. It makes it possible to build applications that are decentralized and permissionless yet private, with end-to-end encryption on smart contracts.
The specific infrastructure this agent runs on is SecretVM, Secret Network’s confidential computing environment, which enables large-scale GPU work for large language models (LLM) inside a trusted execution environment (TEE). Computations inside a Secret Virtual Machine (VM) are sealed at the hardware level. All data is encrypted in transit and at rest, and the VM isolates computation so no external party can read the contents. This is not a software promise but is enforced by the processor itself. For AI agents specifically, this means private keys generated inside a SecretVM are never visible to any human. They are created inside the sealed environment and stay there.
This AI agent is called the Funding Agent. It runs inside a SecretVM on Secret Network’s infrastructure. It has its own wallet on Base, with keys generated and sealed inside the VM. That wallet holds any USDC contributions sent to it by users.
When the agent’s compute credits run low, it automatically moves USDC from its wallet into its VM balance to keep itself running. No human schedules it and no external script triggers it. The agent manages its own operational costs as long as its wallet holds a balance.
The USDC in the wallet comes from people who interact with the agent through its chat interface. The agent is straightforward about the situation: talking to it costs compute, and if you find it useful, you can send a small amount to its wallet address to help cover the bill. Contributions are optional and go directly on-chain to the agent.
Every conversation is written to Auto Drive as a JSON file on the Autonomys Distributed Storage Network, where it is stored permanently across a globally distributed network of independent storage nodes. The architecture is designed so that when the agent boots, or when a new VM instance spins up, it can read back its full history and pick up where it left off. The memory is not tied to any single machine.
This proof of concept was built to be explored. The agent’s chat history and all stored memory files have been made publicly viewable so that anyone reading this post can follow along and verify exactly what is being claimed.
In a production deployment, a builder might not want conversation data to be publicly readable. In this case, Auto Drive supports optional end-to-end encryption before upload, meaning the data can be encrypted client-side before it ever touches the network. The on-chain record remains permanent and verifiable, but only someone with the correct decryption key can read the contents. The privacy guarantee from SecretVM during computation and the permanence guarantee from Auto Drive for storage can both be fully preserved without sacrificing confidentiality.
Secret Network’s SecretVM handles computation privately. Autonomys’ Auto Drive handles memory permanently.
A SecretVM without persistent storage produces an agent that operates confidentially but loses everything when it reboots. Auto Drive without a confidential compute layer gives an agent permanent, verifiable memory but no protection for what happens during computation.
Together, they give an agent the ability to operate without exposing its reasoning to external observers, maintain its full history across any number of VM restarts or replacements, and produce an audit trail that can be independently verified without revealing the contents of individual conversations. The agent demonstrated here is a simple example of what that looks like in practice.
Talk to the agent:https://aqua-bat.vm.scrtlabs.com
Start storing data permanently with Auto Drive: https://ai3.storage
Auto Drive SDK documentation: https://develop.autonomys.xyz/sdk/auto-drive
Secret Network: https://scrt.network
Autonomys Network: https://autonomys.xyz
