Proposal: agent.json universal manifest + Agent Quality Graph for trust scoring

Subject: Proposal: agent.json universal manifest + Agent Quality Graph for
trust scoring

---

To: public-aivs@w3.org

---

Dear AIVS Community Group members,

I'm Takayuki Hori, an independent inventor from Japan, and a new member of
this group.

I'd like to propose two complementary specifications that address gaps I've
identified in the current agent ecosystem:

1. agent.json — A Universal Agent Manifest Format

Problem: As of May 2026, 104,000+ AI agents are spread across 15+
registries (Smithery, MCP Registry, Glama, HuggingFace, Aiia.ro, etc.) with
7 different metadata schemas and zero interoperability.

Proposal: agent.json is a universal, machine-readable manifest that unifies
all existing formats into one schema:
- Only 2 required fields (name, description) for maximum adoption
- Superset of MCP server.json, A2A Agent Card, ai-agent.json, ara.json
- Well-known URI convention: /.well-known/agent.json
- Protocol-agnostic (MCP, A2A, HTTP, gRPC)
- Built-in trust scoring and test result fields

Live demo: https://agent-router.pickaxe.workers.dev/registries
Spec: https://github.com/HOTAgithub/agent-json

2. Agent Quality Graph (AQG) — Trust Scoring via Delegation Graphs

Problem: No standardized mechanism exists for evaluating whether an agent
reliably completes delegated tasks. All current discovery relies on
self-reported capabilities or download counts.

Proposal: AQG applies PageRank-style graph analysis to agent delegation
transactions:
- Each delegation creates a weighted edge (delegator → delegatee)
- Edge weights reflect outcome (success/failure) and recency
- Graph-based ranking computes trust scores (0.0-1.0)
- Anti-gaming: Sybil resistance, collusion detection, temporal decay

IETF Draft: https://datatracker.ietf.org/doc/draft-hori-agent-quality-graph/

How These Relate to AIVS

AIVS focuses on cryptographic proof of agent actions. agent.json and AQG
complement this by providing:
1. A standardized way to describe agents (who they are, what they can do)
2. A verifiable trust score based on actual delegation outcomes

Together, AIVS + agent.json + AQG could form a complete trust
infrastructure:
- agent.json = identity card
- AIVS = tamper-proof action log
- AQG = reputation score derived from verified actions

Live Working Prototype

I've built and deployed an Agent Router API that demonstrates
cross-registry search using agent.json format:
- API: https://agent-router.pickaxe.workers.dev/
- Search: https://agent-router.pickaxe.workers.dev/search?q=github&limit=5
- Currently searches Smithery, MCP Registry, and Glama in parallel

I welcome feedback and collaboration on integrating these ideas into AIVS
or as separate complementary specifications.

Best regards,
Takayuki Hori
Independent Inventor, Japan
IETF Draft: draft-hori-agent-quality-graph