Skip to main content
The foundation of agent sovereignty rests on two critical requirements that must be satisfied before addressing any other challenges.

​
1. Distributed Key Management

A sovereign agent must maintain exclusive control over its private keys and credentials by replicating them across multiple enclave instances running on different types of hardware TEEs. This distribution ensures resilience against both hardware-specific vulnerabilities and cloud provider dependencies. By generating and storing keys within TEEs, the agent can cryptographically prove through remote attestations that no human or external entity has access to its credentials or can manipulate its behavior.

​
2. Secure Updateability

The agent’s codebase must be able to evolve and incorporate new capabilities without compromising the security of its keys or the integrity of its identity. This requires implementing a secure update mechanism that preserves the agent’s state and memory while allowing for the addition of new tools and functionalities. The update process must be decentralized and transparent to prevent any single entity from gaining control over the agent’s evolution.

​
Current Limitations

Modern AI agents have achieved remarkable proficiency in tool usage, decision-making, and environmental interaction. However, they fall short of true autonomy, functioning more as AI experiments with human oversight. In these systems, humans retain ultimate control over the agent’s accounts and actions, fundamentally limiting their sovereignty.

​
Overview of technical Challenges

​
1. Update Control

As agent capabilities and infrastructure evolves over time, a proper architecture for agent updates is required. Any solution must eliminate single points of control while maintaining transparent oversight of the update process. This includes how code changes are proposed, approved, and applied to running agents.

​
2. Memory and State Integrity

Every time an agent updates, its memory and state must transition securely and with proper authentication. The system must prevent arbitrary deletion or modification of memory and state data without proper authorization from a governing body. This requires implementing sophisticated state management protocols that maintain continuity across updates while protecting against unauthorized modifications.

​
3. Tool Interaction Reliability

While agents demonstrate competence in basic tool usage, achieving around 80% success rates with single-call multi-parameter functions, more complex scenarios remain challenging. Success rates for multi-step, multi-turn tool interactions typically fall below 70% for most models, highlighting the need for more sophisticated interaction capabilities.

​
4. World State Verification

Agents require trustworthy mechanisms for observing external world state. Current approaches rely on potentially manipulated API responses, offering no guarantees about information authenticity. Implementing verifiable oracle systems for external data input is critical for enabling autonomous decision-making based on real-world events.

​
5. Web Access Constraints

Agents face significant restrictions on website access, limiting their ability to act autonomously on the internet. This limitation stems from the fundamental challenge of distinguishing between beneficial and malicious agents, or between good and bad principals deploying these agents. Developing reliable authentication mechanisms for legitimate agent interactions remains an open challenge.

​
6. Smart Contract Deployment Safety

While agents can write and deploy smart contract code, these contracts often contain subtle vulnerabilities that are difficult to detect. Any smart contract deployment must undergo rigorous committee scrutiny before putting funds at risk, with comprehensive security checks and risk assessment protocols in place.

​
7. TEE Security and Resilience

There exists an inherent risk of permanently losing access to agent keys and state stored within TEEs. This can occur through various failure modes: compromise of specific TEE technologies, deprecation of cloud provider TEE offerings, or critical software bugs. Implementing robust backup and recovery mechanisms while maintaining security guarantees presents a significant challenge.

​
Future Considerations

A critical next step toward complete agent sovereignty involves placing LLM weights themselves inside TEEs. Current implementations rely on external providers like OpenRouter for LLM API access, introducing trust assumptions around honest execution of inference functions and response routing. Moving the entire inference process inside TEEs would eliminate these dependencies and provide end-to-end sovereignty guarantees.