AgentNet is an open standard and reference architecture for publishing, resolving, and consuming authoritative machine-readable data for AI systems.
It introduces a machine-first web layer built on JSON-LD, enabling AI agents to retrieve grounded, provenance-aware, policy-constrained information directly from publishers—without scraping, hallucination, or opaque intermediaries.
Modern AI systems rely heavily on probabilistic inference over untrusted or ambiguous data. This creates systemic problems:
- Hallucinations and unverifiable claims
- Loss of source authority and provenance
- No enforceable usage terms or attribution
- Inefficient retrieval (token waste, over-contextualization)
- No economic or governance layer for machine access
AgentNet addresses this by defining a federated, standards-based system where:
- Publishers declare authoritative facts in machine-readable form
- AI agents retrieve only what is explicitly published and permitted
- Provenance, versioning, and usage constraints are first-class
- Retrieval is deterministic, efficient, and auditable
AgentNet is composed of four foundational elements:
-
Nodes
Publishers (organizations, creators, datasets, systems) that expose authoritative data. -
Capsules
JSON-LD documents that encapsulate facts, metadata, provenance, and usage constraints. -
Resolvers
Federated services that locate, validate, and return the appropriate capsule(s) for a given request. -
Registrars
Services that manage identity, registration, and trust anchors for nodes and resolvers.
Together, these form a machine-centric information layer optimized for AI retrieval and reasoning.
This repository is the public front door for AgentNet.
It contains:
- The authoritative AgentNet Standards (ANS)
- Conceptual and architectural documentation
- Reference materials and examples
- Governance and roadmap artifacts
This repo is intentionally communicative and stable.
Active implementation code lives in related repositories.
AgentNet/
├─ README.md ← You are here
├─ standards/ ← AgentNet Standards (ANS)
│ └─ ANS-Core-v2.0.docx
├─ docs/ ← Conceptual & architectural documentation
├─ examples/ ← Example capsules, requests, and responses
├─ press/ ← One-pagers, fact sheets, media assets
├─ GOVERNANCE.md ← Governance model (bootstrap)
├─ ROADMAP.md ← Public roadmap
├─ LICENSE
├─ SECURITY.md
├─ CONTRIBUTING.md
└─ CODE_OF_CONDUCT.md
This structure is designed to support multiple audiences simultaneously:
- Standards readers should begin in /standards
- Architects and integrators should explore /docs
- Developers and evaluators should inspect /examples
- Press and partners may reference /press
- Contributors and stewards should review governance and policy documents at the root
The layout is intentionally conservative and explicit to ensure long-term clarity, stability, and ease of navigation.
The AgentNet Standards (ANS) define the normative rules of the AgentNet ecosystem.
- Architecture and terminology
- Capsule structure and requirements
- Resolver behavior and selection logic
- Provenance, trust, and governance principles
- Compliance and interoperability expectations
📄 Start here:
/standards/ANS-Core-v2.0.docx
ANS Core v2.0 is final and effective as of 2026-01-01, superseding all prior drafts.
If you are new to AgentNet:
- Read ANS Core v2.0 (overview sections first)
- Review
/docs/overview.md - Inspect example capsules in
/examples/capsules - Explore the reference resolver repository (linked below)
You do not need to run infrastructure to understand the system.
AgentNet is a standard, not a single product.
However, reference implementations exist to demonstrate compliance and feasibility.
Related repositories:
- Reference Resolver: implements capsule resolution, trust validation, and policy selection
- Capsule Generator (experimental): demonstrates automated capsule creation
- Validator: checks capsule conformance against ANS
Links are provided in /docs/implementations.md.
- Standards: Stable (ANS Core v2.0)
- Architecture: Proven and tested
- Reference Resolver: Functional and compliant
- Ecosystem: Early, intentionally open
AgentNet is designed for incremental adoption—publishers and consumers can start small and expand.
AgentNet is governed as an open, federated standard.
- No single resolver, registrar, or operator is required
- Governance roles are defined but initially unfilled
- Emergency changes require post-hoc disclosure
- Long-term stewardship is designed to be multi-stakeholder
See GOVERNANCE.md for details.
AgentNet is relevant to:
- AI platform builders and model providers
- Data publishers and content owners
- Enterprises concerned with AI grounding and provenance
- Infrastructure and tooling companies
- Regulators, standards bodies, and policy groups
- Investors evaluating foundational AI infrastructure
Contributions are welcome—but standards quality matters.
Please read:
CONTRIBUTING.mdCODE_OF_CONDUCT.md
Standards changes follow a structured review process.
If you discover a vulnerability or integrity issue related to AgentNet reference implementations, please follow the process in SECURITY.md.
Unless otherwise noted, this repository is licensed under the terms in LICENSE.
AgentNet is intentionally conservative, explicit, and boring where it matters.
That discipline is what enables trust, interoperability, and long-term adoption in a machine-centric web.
If you are building AI systems that need to know where their facts come from—you are in the right place.