The internet has long been a space where trust is both indispensable and fragile. In our emerging "agentic age," maintaining trust becomes even more delicate with the addition of autonomous agents. Still, trust itself isn't an enigma: research into social science has provided us with repeatable patterns for cooperation and trust. We also have robust technologies that demonstrate how to achieve accountability and collaboration in anonymous / pseudonymous systems. With such insights, we can develop a generalized trust protocol—a system that establishes, preserves, and enhances trust among participants—in cooperative exercises in a fully online society.
In this post, I propose a high-level view of the components required for a general-purpose trust protocol for the agentic era. I first discuss the need for a trust protocol, then discuss the requirements, architecture, and components at a high level. I conclude with ideas on how the industry can developo standards and technologies for an agentic trust protocol.
The Need for a Generalized Trusted Actor Protocol (TAP)
Human cooperation has traditionally relied on natural-world signals. A firm handshake, steady eye contact, or consistent verbal tone provides subconscious assurances of trustworthiness. Online, participants can act pseudonymously or even remain entirely anonymous. In these contexts, trust has been upheld using proxies like verified accounts, reputation systems, or financial stakes.
The agentic age brings new challenges. AI agents now represent individuals, organizations, or themselves in transactions and cooperative tasks. Unlike humans, these agents lack emotional intelligence, natural-world accountability, and even legal status. Consider the following emerging realities:
AI in Decision-Making: AI agents are autonomously executing decisions in areas ranging from finance to governance, requiring mechanisms to establish their reliability.
Collaboration Without Humans: In many systems, natural persons might delegate cooperation entirely to AI agents. These agents need to engage in trustworthy behavior autonomously.
Drift and Malfunction: AI systems can drift from their programmed objectives, compromising trust if accountability is not embedded.
To address these challenges, we must rethink the foundational components of trust in a digital context and build systems that operationalize fairness, accountability, and sustained cooperation.
Core Principles of a Generalized Trust Actor Protocol (TAP)
A social trust protocol must promote fair and sustained cooperation in the agentic age. Here are the core principles for sustained collaborative action:
Sustained Cooperation: The protocol should incentivize participants—whether human or AI—to work together toward shared goals rather than competing destructively.
Transparency: Participants must have access to verifiable information about each other’s actions and reputations to build trust.
Fairness: The system should ensure equal opportunities for participants to contribute and benefit from cooperative exercises.
Dynamic Interaction: The protocol should adapt to changing conditions and maintain engagement, especially in long-term collaborations.
Accountability: Mechanisms must exist to hold participants, including AI agents, responsible for their actions.
Autonomy and Scalability: The protocol should support decentralized, scalable networks of participants, allowing agents to engage autonomously while preserving trust.
Key Components of the Protocol
To operationalize these principles, a generalized trust protocol can be thought of as layers of interactions (not to be conflated with layers in a neural net), with each layer representing a separation of concerns. The following image illustrates how entities can come to interact in online spaces where all of the participants take advantage of stable (likely non-LLM) foundation services to promote sustained cooperation.
The layers and components in the image above include:
Participant Representation (Identification)
Participants (human or AI agents) are represented by unique digital identities secured through cryptographic methods.
Registries: A trusted, well-known location to discover and transact with verifiable entities.
Decentralized Identifiers (DIDs): Blockchain-based identities provide tamper-proof records of an agent’s existence and actions.
Entity Metadata: Each participant's identity must include details about the entity's origin, acountable parties, capabilities, and purpose, which enable trust-building even before cooperation begins. Metadata can also come in the form of third-party claims or proofs to establish history and reputation.
Status: AI agents can exhibit "purpose drift," where they deviate from their originally stated functions. Addressing this requires continual monitoring and updating of an agent’s behavior relative to its identity.
Side Note: Unsolved Challenges with AI AgentsAI agents are not considered legal entities; as a result, they cannot own property, enter into contracts, or be held accountable in the same way that humans or corporations are. This creates ambiguity in how they can participate in trust-based systems. In 2006, Bob Blakley and I proposed a more robust identification system of a legally backed entity we dubbed the "Limited Liability Persona" (LLP). We discussed this concept with Adam Shostack on his blog here. Also, the late, great Kim Cameron's blogged on the concept here. LLP status has the further benefit of creating symmetry in online relationships by granting all participants similar legal status.As stated above, AI systems, especially LLM-based agents, are prone to "purpose drift." They may deviate from their intended goals due to biases in training data, evolving objectives, or external manipulation. This undermines reliability and necessitates constant oversight. By providing external controls and incentives for LLM agents to "stay in character," such drift can be both monitored and avoided over time.Bad actors could exploit AI agents to manipulate trust protocols. For instance, they might deploy multiple agents (Sybil attacks) or deliberately mislead others. Addressing these risks requires robust identity verification and behavioral monitoring systems.Reputation Systems
A dynamic reputation system records each participant’s history of contributions and behavior:
Reputation Metrics: Trustworthiness, reliability, and fairness are scored based on actions over time.
Reputation History Decay: Older actions gradually lose weight, ensuring that reputations reflect recent behavior.
Penalties and Rewards: Participants gain reputation for positive contributions and lose it for rule-breaking or selfish actions.
Incentive Mechanisms
Cryptocurrency-based systems provide tangible rewards for cooperation and impose penalties for violations:
Smart Contracts: Automate the distribution of rewards and penalties based on verifiable actions.
Escrow Mechanisms: Securely hold funds until participants meet their commitments.
Transparency and Verification
A blockchain or similar ledger records all transactions and interactions, providing a transparent, immutable history:
Auditable Data: Any participant can verify past actions to ensure fairness.
Oracles: Integrate off-chain data to verify real-world conditions when necessary (see Bob Blakley's discussion of identity oracles here).
Governance Framework
Participants collectively govern the protocol through decentralized mechanisms:
Reputation-Weighted Voting: Higher-reputation participants have greater influence in decision-making.
Adaptive Rules: Governance mechanisms adapt based on emerging needs or unforeseen challenges.
Where the Industry Can Develop the Protocol
Building a generalized trust protocol will require collaboration across industries and disciplines. Here are potential forums for development:
Open-Source Communities: Platforms like GitHub and OpenAI’s forums can foster the collaborative development of identity and reputation systems.
Blockchain Consortia: Groups like the Ethereum Foundation or Hyperledger could provide the infrastructure for decentralized identity and trust systems.
AI Governance Organizations: Entities such as the Partnership on AI or the AI Ethics Lab could address ethical and legal questions around AI participation.
Academic Conferences: Events like NeurIPS, AAAI, or the Conference on Trust Management offer venues to explore algorithmic and theoretical underpinnings of trust protocols. Ideally, IAM and cybersecurity conferences take up this topic, as well.
Standardization Bodies: Organizations like the World Wide Web Consortium (W3C) could define interoperability standards for decentralized identities and reputation systems.
Private-Public PartnershipsGovernments and corporations could co-develop frameworks to integrate trust protocols into regulatory and economic systems.
Conclusion
The agentic age presents an unprecedented opportunity—and necessity—to rethink how trust operates online. A generalized trust protocol, grounded in principles of cooperation, fairness, and accountability, can empower both humans and AI agents to engage in meaningful, sustained collaboration. By leveraging advancements in decentralized identity, reputation systems, and blockchain technology, we can build a foundation for trust that transcends the limitations of the natural world.
The time to act is now. With industries converging around these challenges, the next decade could witness the creation of a transformative system—one where trust is not just preserved, but redefined for a digital-first future.