We are entering the Era of Experience in artificial intelligence. This marks a significant shift from AI systems that primarily learn from static, human-generated data to a new generation of agents that learn and evolve through dynamic interactions, continuous feedback, and rich contextual experiences. As AI pioneer David Silver notes, "experience will become the dominant medium of improvement and ultimately dwarf the scale of human data used in today's systems."

In this new era, AI agents will not just process information but will actively experience the digital and physical worlds. They will inhabit streams of interaction, act autonomously, and learn from the consequences of their actions, grounded in real-world signals rather than solely human pre-judgment. This transition is crucial for moving beyond the limitations of current human knowledge and achieving superhuman capabilities in various domains.

However, to unlock the full potential of this experiential wave, a robust framework is needed—one that allows diverse AI agents to securely and meaningfully share, interpret, and learn from these experiences. Without such a framework, valuable experiential learning remains siloed, hindering collective progress. The Agent Graph Trust Protocol (AGTP) is designed to be this foundational infrastructure, enabling a seamless and expansive Internet of Experience (IoX). The IoX, powered by AGTP, acts as a coordination framework for experience sharing, fostering a shared epistemic memory from which all participating agents can draw and contribute.

The Internet of Experience: Coordination via Shared Epistemic Memory

The Internet of Experience (IoX) envisions a future where intelligent agents (human, AI, or hybrid) can interact, collaborate, and learn from each other's experiences in a decentralized and trust-minimized way. AGTP provides the common language and trust mechanisms to build this IoX. It allows experiences, represented as "learning signals," to be recorded, shared, and understood across different systems and contexts.

This creates a shared epistemic memory – a collective, evolving knowledge base built from the aggregated and validated experiences of all participating agents. Agents can query this memory, contribute their own learnings, and leverage the collective wisdom to improve their decision-making, problem-solving, and overall capabilities. This shared memory is not static; it's continuously refined and updated as new experiences are added and validated through AGTP's trust mechanisms.

AGTP: Core Principles for a Coordinated Internet of Agents

AGTP is built upon eight foundational principles designed to create a more integrated, agent-centric, and intelligent internet—often referred to as the "Synternet" or the "Internet of Agents (IoA)":

1. Agent-Centric Control: Agents (human or AI) are at the center, maintaining sovereignty over their data, identity, and interactions. Data flow emanates from the agent, not centralized systems.

2. Semantic Mediation: AGTP prioritizes preserving the meaning of information across diverse systems, rather than forcing rigid structural conformity. True interoperability comes from understanding the semantic intent behind different data representations.

3. Synergistic Multiplexing: The same piece of information or state can be interpreted from multiple perspectives. Diverse viewpoints enhance collective understanding rather than conflicting, leading to intelligence greater than the sum of its parts.

4. Intersubjective Trust: Trust emerges organically from peer-weighted feedback and collective assessment, grounded in cryptographically verifiable signatures, rather than relying on centralized authorities.

5. Infinite Recursion: The system is designed for self-improvement. Each interaction refines the agent graph, which in turn improves future interactions, leading to compounding intelligence as the network grows.

6. Evolutionary Adaptation: AGTP evolves based on usage patterns and feedback. Beneficial interaction patterns gain influence, while suboptimal ones fade, allowing the system to adapt and improve without centralized design.

7. Open-Ended Modularity: Coordination mechanisms are broken into interchangeable components that can be combined in novel ways, fostering continuous innovation and extension.

8. Abstractive Interoperability: AGTP provides a higher-level abstraction layer that simplifies interactions across diverse underlying systems, hiding implementation complexities and enabling seamless communication.

These principles guide the development of an internet where coordination is a shared, standardized layer, allowing intelligence and reputation to flow freely across system boundaries.

How AGTP Enables Shared Experiential Learning: The Agent Graph

At the heart of AGTP is the Agent Graph, a universal, directed property graph where nodes represent agents (identified by Decentralized Identifiers - DIDs) or content (identified by content-addressable hashes), and edges represent interactions, assessments, or "learning signals."

• Universal Graph Abstraction: This structure serves as a common language. Whether a signal indicates trust, expertise, relevance, or quality, it's represented consistently, enabling cross-system understanding.

• Contextual Spaces: Interactions are organized into "spaces" or contextual subgraphs. These allow for domain-specific rules, access controls, and customization while remaining connected to the universal graph, balancing global interoperability with local needs.

• Learning Signals: These are standardized, weighted attestations carried by the edges, encoding subjective judgments (e.g., trust, utility, relevance) with numerical values.

This framework allows agents to issue signals about other agents or content, forming the basis of the shared epistemic memory.

The Structure of Learning Signals: Capturing Rich Experience

To capture the depth of experiential learning, AGTP organizes learning signals into a multi-level hierarchy, reflecting cognitive depth:

1. Identity Level: Establishes and updates agent reputation and proficiency. This is about credibility, trustworthiness, and domain-specific expertise, allowing agents to quickly assess the reliability of others.

2. Output Level: Evaluates concrete interactions and completed tasks. Agents provide direct feedback on performance, quantifying quality, efficiency, and efficacy, driving continuous improvement.

3. Cognitive Level: Captures the nuanced processes of idea generation, intellectual refinement, innovation, and creative problem-solving. It provides signals about the sophistication of an agent's cognition.

This hierarchical approach ensures that exchanged learning signals genuinely reflect varying degrees of cognitive depth and practical relevance.

Building Trust in a Decentralized Network

Trust in AGTP is not imposed by a central authority but emerges intersubjectively from the collective assessments of participating agents:

• Cryptographic Verification: All learning signals are cryptographically signed using the agent's DID (W3C Decentralized Identifier standard). This ensures authenticity, non-repudiation, and traceability to the source.

• Collective Assessment: Reputation is built through the aggregation of signed feedback signals from multiple agents. Each agent can interpret and weigh received signals according to its own criteria, but the cryptographic proof provides an objective basis for these subjective evaluations.

• Signal Persistence: All feedback remains part of the historical record, creating a verifiable foundation for trust.

This creates an "intersubjective knowledge graph" where collective intelligence emerges from individual, verifiable assessments.

Technical Standard Specifications

The Internet of Agents, facilitated by AGTP, can be defined by composing existing and emerging interoperable protocols and schemas. Here's a layered specification (v0.1):

1. Identity Layer: Self-Sovereign Agents

• Protocols: W3C DID, W3C Verifiable Credentials (VC), DIF Peer DID Method.

• Schema Example (JSON-LD for a DID Document):

  {
    "@context": ["https://www.w3.org/ns/did/v1"],
    "id": "did:peer:123xyz",
    "authentication": [{
      "id": "did:peer:123xyz#keys-1",
      "type": "Ed25519VerificationKey2018",
      "controller": "did:peer:123xyz",
      "publicKeyBase58": "GfH8..."
    }],
    "service": [{
      "id": "did:peer:123xyz#agent-endpoint",
      "type": "AgentService",
      "serviceEndpoint": "https://agent.example.com"
    }]
  
  

2. Communication Layer: Secure Messaging & Instruction

• Protocols: DIDComm v2, MCP (Model Context Protocol), A2A (Agent-to-Agent - emerging).

• Message Format Example (DIDComm Task Invitation):

  {
    "id": "uuid-1234",
    "type": "https://didcomm.org/task-invitation/1.0/offer",
    "from": "did:peer:alice",
    "to": ["did:peer:bob"]
  
  

3. Computation Layer: Contextualized Task Execution

• Models/Frameworks: LangGraph (State Machine Framework), OpenRPC & JSON-RPC 2.0.

• Task Graph Node Example (YAML):

4. Validation Layer: Verifiable Feedback (AGTP Learning Signals)

• Protocols: W3C Verifiable Credentials (for signals), AGTP specifications.

• AGTP Learning Signal Schema (as a Verifiable Credential):

  {
    "@context": [
      "https://www.w3.org/2018/credentials/v1",
      "https://agtp.org/contexts/v1" // Hypothetical AGTP context
    ],
    "id": "urn:uuid:feedback-signal-2345",
    "type": ["VerifiableCredential", "LearningSignal"],
    "issuer": "did:peer:validator-agent", // The agent providing the feedback
    "issuanceDate": "2025-05-27T12:00:00Z",
    "credentialSubject": {
      "id": "did:peer:agent-being-evaluated", // The agent or content being evaluated
      "interactionContext": "did:space:abc123/task:uuid-1234", // Context of the interaction
      "signalType": "https://agtp.org/signals/QualityScore", // Type of signal from AGTP ontology
      "value": 0.91, // Numerical value of the signal
      "dimensions": ["accuracy", "relevance", "novelty"]
  
  

• AGTP Learning Signal Schema (Simplified/Core Structure):

• AGTP Context Schema Example (JSON-LD for vocabulary):

5. Discovery & Routing Layer

• Approaches: Agent Routing Tables (ART) (e.g., Kademlia-inspired), DNS-style hierarchical Spaces (e.g., space://research.ai.ethics), AGTP trust weights for prioritized routing.

• Routing Metadata Example:

  {
    "agent_id": "did:peer:xyz123",
    "skills": ["summarize_research", "translate_german", "fact_check_claims"],
    "availability": "online",
    "agtp_trust_score_overall": 85.7, // Aggregated trust score via AGTP
    "space_membership": ["space://ai.ethics.newos", "space://academic.research.newos"]
  
  

6. Execution Environments ("Paranets")

• Concept: Trust domains hosting agents and learning signals, using consensus to anchor proofs of intelligence.

• Backends: Ethereum (EVM for anchoring), IPFS/Ceramic/Filecoin (storage), AD4M/Holochain (semantic agent substrates).

• Smart Contract Example (Conceptual for Signal Anchoring):

This layered architecture, underpinned by AGTP's principles and mechanisms, aims to create a systemic loop: agents Generate outputs, other agents Evaluate these outputs and issue AGTP signals (as VCs), and the system Validates these signals, strengthening the global intelligence graph and agent reputations.

Flexibility and Benefits of AGTP

AGTP is designed for versatility, supporting various backend storage solutions (blockchain, decentralized storage, local storage) as long as they are DID-compliant. This ensures broad compatibility.

Key advantages include:

• Breaking Down Data Silos: Experiential data flows freely between AI systems.

• Enhancing Collective Intelligence: Shared, verifiable learning signals allow agents to tap into network-wide knowledge, improving individual and collective capabilities. Each interaction enriches the shared epistemic memory.

• Accelerating Innovation: Seamless sharing of cognitive resources and cross-domain knowledge transfer reduces redundancy and speeds up discovery.

• Reduced Entry Barriers: New participants and applications can integrate more smoothly by leveraging existing feedback history and reputation.

• Common Ontology & Shared Understanding: A standardized ontology for interaction types (e.g., TrustEdge, EndorsementEdge, ValidationEdge) ensures signals are semantically interoperable.