RFP-9: ICP as Infra for AI Agent Frameworks

[jekennedy]

Overview

The Internet Computer (ICP) provides a decentralized, secure, and interoperable platform that can be used by AI Agent Frameworks to build autonomous, tamper-proof systems. Canister smart contracts on the ICP serve as powerful backbones for AI agents, enabling features such as persistent memory, decentralized governance, trustless value transfer, and integration with multi-chain ecosystems through Chain Fusion.

By utilizing the ICP’s unique capabilities (see References section), developers can create AI agents that are more secure, composable, and economically autonomous than traditional centralized solutions. Additionally, developers can leverage canisters to build, deploy, and run AI agents natively on the ICP, similar to how alice.fun has implemented agent execution within ICP canisters, ensuring high scalability, efficiency, and on-chain execution integrity.

Security Considerations

AI agents interacting with the ICP must ensure the security and integrity of their operations. Proposals should address the following considerations:

• Canister-Based Deployment for On-Chain Security: AI agents deployed directly within ICP canisters benefit from built-in security mechanisms such as tamper-proof execution, cryptographic guarantees, and on-chain state persistence. Running agents within canisters eliminates reliance on external infrastructure, reducing vulnerabilities related to centralization and third-party breaches. Additionally, canister-controlled access permissions ensure that AI agents operate within strictly defined parameters, preventing unauthorized modifications or malicious overrides.

• Trusted Execution Environments (TEEs): Running an agent in a TEE provides certain guarantees about the confidentiality and integrity of its computations, ensuring that the agent’s code and data remain secure even when executed on potentially untrusted hardware or infrastructure. Existing tools are available in the ICP ecosystem, such as the work already done by LDC Labs:

  • IC-TEE: Make Trusted Execution Environments (TEEs) work with the IC.
  • IC-COSE: A decentralized Configuration service with Signing and Encryption on the IC.

• Secure Communication: Ensure end-to-end encryption (e.g., TLS) for all interactions between the agent and the IC.

Request for Proposal (RFP) Categories

Integrations with Existing AI Agent Frameworks

We encourage proposals that integrate with established AI Agent Frameworks to provide new plugins that utilize key features of the ICP (see References section below) for AI Agent infrastructure. Examples of existing Web3 AI Agent Frameworks:

• ZerePy
• Eliza
• HeyAnon

Integrations with the LLM Canister

The LLM Canister enables the integration of LLMs into ICP applications. With just a few lines of code in Rust or Motoko, developers can prompt LLMs within their canisters.

We encourage proposals that integrate the LLM Canister to build on-chain AI agents for:

• AI-Driven DAOs: AI agents can leverage the ICP to operate as decentralized autonomous organizations (DAOs). By using the IC’s governance tools, an agent could:

  • Propose and vote on operational decisions autonomously.
  • Collaborate with other agents and users in a transparent governance framework.
  • Employ the Service Nervous System (SNS) for funding, community engagement, and operational scalability.

• Wallet/Identity Functionality: AI agents can use the ICP for wallet and identity services, allowing it to manage digital assets, linked with Chain Fusion to interact across multiple chains. Some example capabilities:

  • Automating on-chain transactions with built-in thresholds and multi-signature capabilities.
  • Leveraging Internet Identity for secure delegation, allowing agents to act on behalf of users in a trust-minimized way (e.g., executing trades, managing funds, or interacting with dApps based on pre-set permissions).
  • Acting as escrow managers in multi-party transactions, holding assets securely on behalf of users or other agents until conditions are met.

• DeFi Interactions: AI agents can interact with DEXs and other DeFi protocols directly on the ICP, within ICP-specific ecosystems (e.g., Bitfinity), or across other blockchain ecosystems using Chain Fusion.

• Verifiable Credentials: AI agents can leverage verifiable credentials to prove identity, reputation, or authorization, enabling trust-based interactions in DeFi, DAOs, marketplaces, and Web2/3 services. Some example capabilities:

  • As an Issuer:

    • DeFi Credit Scoring AI: An AI agent analyzes on-chain lending activity and issues a “VerifiedBorrower” credential proving that a user has repaid past loans without defaults.
    • In a DAO or Web3 marketplace, an AI agent could aggregate on-chain activity, github contribution history, and governance participation, issuing a “VerifiedContributor” credential.

  • As a Subject:

    • Autonomous DeFi Trader AI: An AI agent requests and presents a “VerifiedTrader” credential proving it has a history of successful arbitrage trades before being allowed to manage funds on behalf of users.
    • AI DAO Delegate: A DAO governance body issues a “DelegatedVotingPower” credential to an AI agent, proving it has been authorized to cast votes on behalf of a human or another entity.

  • As a Relying Party:

    • A DeFi compliance AI verifies the credential:
      • “VerifiedResident” before allowing users to interact with on-chain DEXs or staking services.
      • “VerifiedAdult” before allowing users to interact with the Agent.

Extending the Anda Agent Framework

We are looking for proposals to expand the capabilities of the Anda Framework (an existing ICP-based agent framework) by adding features such as those listed below. Grant recipients may also be eligible for funding from the Community Incentive Fund managed by the ICPanda DAO.

• Use Internet Identity for delegation to Anda agents.
• Improved functionality of the Anda engine and core.
• Enabling any of the previously mentioned Novel Protocols and Applications, in the context of the Anda Framework.

Proposals should also consider including in their overall solution the creation of new Anda tools, to enable its agents to directly interact with ICP canisters (note: existing tools are limited to balance queries and token transfers).

AI Agents as Canister-Based Autonomous Systems

AI agents can be deployed directly within ICP canisters, allowing them to function as standalone, self-governing digital entities. These agents can:

• Manage smart contract-based funds autonomously.
• Make decisions based on input from on-chain data, off-chain APIs, and AI models.
• Interact with external data sources via HTTPS outcalls.

Tutorials and Well-Documented Sample Applications

We are looking for high-quality, well-documented resources to help new developers integrate the ICP into AI agent frameworks. Examples include:

• Tutorials: Step-by-step guides for building an AI agent that uses the ICP as infrastructure.
• Sample Applications: Real-world use cases showcasing how AI agents can:
  • interact with the ICP’s DeFi ecosystem
  • leverage Chain Fusion for multi-chain interoperability
  • integrate with ICP’s identity and wallet solutions.

How to apply?

Please submit your application at https://dfinity.org/grants, mentioning RFP-9 in the application.

References

• Identity Delegation with Internet Identity
• Chain Fusion Technology for direct integration with Bitcoin, Ethereum, and among others
• Verifiable Credentials for validating an individual's identity or qualifications
• vetKeys to facilitate onchain encryption (an upcoming feature)
• Timers for scheduled onchain tasks
• HTTPS outcalls for interacting with external APIs
• Reverse Gas Model
• Internet Computer Developer Documentation

AI Agent References

• LLM Canister
• Pickpump ICP Plugin that enables Eliza agents to create meme coins on ICP by connecting to the Pickpump canister
• Dollar cost average (DCA) agent example
• llama.cpp for the Internet Computer
• Open-source repository for AI deployment on ICP
• [Eliza AI Agent Framework](https://elizaos.github.io/eliza/)
• [alice.fun](http://alice.fun)

[tiagoicp]

Thanks Justin, this is very useful. Will share with the DevRels of the Hubs so we can share with participants of Bootcamps / Hackathons of our HUBs.