Autonomify

AI agents are becoming a popular interface for Web3, but they cannot yet be trusted to execute on-chain transactions safely. Granting agents execution authority exposes systems to risks such as prompt injection, memory poisoning, and model manipulation. Introducing human supervision may mitigate these risks but breaks autonomy and prevents true automation. This creates an execution dilemma for autonomous systems.

Autonomify solves this by separating proposal from authorization and execution.

Autonomify enables on-demand AI agents with on-chain capabilities while ensuring they never control execution directly. Using CRE, agent-proposed transactions are routed through a secure execution workflow that includes attestable zero-knowledge policy validation and transaction simulation via Tenderly before approved actions are forwarded to a delegated executor contract.

• Agents propose transactions but never hold keys or signers (true AI-assistance architecture)
• ZK proofs verify policy compliance without revealing the policy itself
• CRE orchestrates the secure workflow across the enclave, simulation, and network.
• Delegated execution via ERC-7710 lets the user's smart account remain msg.sender

• User policies are stored inside an AWS Nitro Enclave, where a Noir zero-knowledge circuit generates proofs verifying that a proposed transaction satisfies the user’s constraints without revealing the policy itself. The enclave image includes the Barretenberg proving library and runs in isolation.

• CRE coordinates the workflow by receiving agent execution requests, requesting ZK proofs from the enclave, and forwarding verified reports to an EVM-compatible executor contract written in Solidity. The executor validates the proof and executes the transaction on the user’s smart account using delegated execution permissions via the MetaMask Delegation Framework (ERC-7710), enabling policy-controlled execution on the user's behalf without requiring access to the private key. The demo executor contract is on Base Sepolia.

• Tenderly provides infrastructure for contract interaction and execution verification. CRE normally reports only the outer transaction status once a transaction is mined, which may hide failures deep in the call stack. To provide accurate feedback for autonomous agents, Autonomify uses Tenderly for ABI retrieval, pre-flight simulation on Virtual TestNets, and post-execution call tracing via debug_traceTransaction to verify the true outcome of internal smart contract execution.

• The AI agent system is built using Vercel’s AI SDK. When a contract is added to an agent, the LLM analyzes the contract ABI fetched from Tenderly, annotates available functions, performs a quick security review, and exports the result in a universal OpenAI-compatible tool format so agents can use them as callable on-chain actions. The frontend also provides a launch interface for non-technical users to deploy and use on-demand AI agents from platforms such as Telegram and Discord.

Autonomify Code References

• Chainlink CRE- docs/CRE.md
• Tenderly - docs/TENDERLY.md
• ERC-7710 Delegation- docs/DELEGATION.md
• ZK Proofs (Noir)- docs/ZK.md