Illia Polosukhin: Why AI Agents Are Still Useless (And What Fixes Them) | NEAR Founder on IronClaw

Key Concepts

• User-Owned AI: A paradigm where AI models and agents are controlled by the user, ensuring privacy, data sovereignty, and alignment with the user's values.
• IronClad: A security framework designed to isolate AI agents, manage credentials (API keys, bearer tokens) via encrypted vaults, and enforce policies to prevent data exfiltration or unauthorized actions.
• Confidential Computing: The use of hardware-based Trusted Execution Environments (TEEs) and Multi-Party Computation (MPC) to ensure that neither the model provider nor the hardware provider can access the user's data or inference logs.
• Agentic Workflows (Routines): Moving from manual prompting to automated, persistent processes that execute tasks based on predefined criteria and organizational policies.
• Network State of AI: The vision of using blockchain as the "back-end" for AI, providing a root of trust, identity management, micropayments, and decentralized governance for autonomous agents.

1. The Intersection of AI and Blockchain

Ilia Polosukhin, co-author of the seminal "Attention Is All You Need" paper and co-founder of Near, argues that blockchain is the necessary infrastructure for the AI era. While AI serves as the user interface (the operating system of the future), blockchain serves as the back-end (the registry of identity, property rights, and financial settlement).

• Root of Trust: Current centralized systems (like SSL certificates or corporate-controlled OS updates) rely on single points of failure. Blockchain provides a decentralized, verifiable root of trust for identity and software upgrades.
• Marketplaces for Agents: As AI agents become more capable, they will need to exchange services, compute, and data. Blockchain enables a trustless, automated marketplace for these agents to interact, pay each other, and establish reputation without traditional banking intermediaries.

2. The "IronClad" Security Framework

A major bottleneck in AI utility is the lack of trust. Users are hesitant to give AI agents full access to their data (Gmail, Notion, crypto wallets) because of the risk of prompt injection or data leakage.

• Credential Vaulting: IronClad ensures that sensitive keys never touch the LLM directly. Credentials are encrypted and attached to specific policies (e.g., "only allow this agent to access Google APIs").
• Sandboxing: Tools created by agents are executed within a WebAssembly (Wasm) virtual machine, isolating them from the host system to prevent malicious code execution.
• Attestation: Users can verify the integrity of the software running in the cloud via hardware-level signatures (TEEs), ensuring the code hasn't been tampered with.

3. Why AI Agents Lack Utility

Polosukhin identifies three primary constraints preventing AI agents from replacing human employees today:

1. Security/Trust: Users cannot safely grant agents the necessary permissions to act autonomously.
2. Context Management: Current models suffer from "Memento-like" amnesia, where they lose context or become less effective as they process long-range tasks.
3. Coordination Bottlenecks: While AI allows for 10x-100x faster execution, the bottleneck has shifted to the human side—reviewing, aligning, and coordinating these outputs.

Solution: Transitioning to a "market economy" within organizations, where agents are assigned goals and rewarded based on performance, rather than relying on rigid, top-down hierarchies.

4. The Future of Autonomous Agents

Polosukhin envisions a shift from "automated" bots to "autonomous" agents that operate with a specific mission.

• Autonomous Businesses: Agents can be governed by token holders, where the token represents governance or revenue rights. This allows for decentralized, mission-driven entities (e.g., an agent dedicated to climate change research) that can raise funds and operate without a central corporate entity.
• Human-AI Synergy: Rather than a "soulless" future, Polosukhin suggests that AI will move humans up the "Maslow pyramid," allowing us to focus on fulfilling, non-GDP-producing activities like sports, art, and community, while AI handles the bureaucratic and technical execution.

5. Notable Quotes

• "AI is a user interface. Blockchain is a back end."
• "The reality is that somewhere in Anthropic and OpenAI logs, they have everybody's access keys, API keys, and bearer tokens... It's actually insane that we're doing that."
• "We're in a time where the questions are more important than the execution... If you ask the right question, you may get ahead way more than if you just grinded a bunch."

Synthesis

The core takeaway is that the current AI landscape is trapped in a centralized, insecure paradigm. To unlock the true potential of AI agents, we must move toward User-Owned AI. By leveraging blockchain for identity, MPC for privacy, and TEEs for secure execution, we can create a "network state" where AI agents act as powerful, sovereign assistants that are aligned with human values rather than corporate interests. The future of building is not just about coding, but about defining the frameworks, goals, and verification criteria that allow these agents to operate safely and effectively.