| # AgenThink Mesh — Protocol Specification |
| ## NAN-RELAY/1.0 · NAN-HS/1.0 · Grounding Gate/1.0 |
|
|
| **Status:** Draft v0.1 |
| **Authors:** AgenThink Engineering |
| **Date:** March 2026 |
|
|
| --- |
|
|
| ## 1. Overview |
|
|
| The **National Agentic Network (NAN) Protocol** defines how AI agents discover, authenticate, and communicate with each other via the AgenThink Mesh orchestration layer. |
|
|
| It answers three questions every multi-agent system must answer: |
|
|
| 1. **Where is the right agent?** → Universal Agent Directory (UAD) |
| 2. **What does the next agent need to know?** → State-Relay Protocol |
| 3. **Is the agent's output trustworthy?** → Grounding Gate |
|
|
| --- |
|
|
| ## 2. NAN Address Scheme |
|
|
| Every agent in the Mesh has a globally unique NAN address: |
|
|
| ``` |
| nan://<country>/<domain>/<name>/<version> |
| |
| Examples: |
| nan://kw/finance/etf-allocator/2.1 |
| nan://global/research/web-researcher/2.1 |
| nan://ae/legal/adgm-compliance/1.0 |
| ``` |
|
|
| **Fields:** |
| - `country` — ISO 3166-1 alpha-2 country code, or `global` |
| - `domain` — functional category (`finance`, `legal`, `research`, `strategy`, `compliance`, `meta`) |
| - `name` — kebab-case agent identifier |
| - `version` — semantic version (major.minor) |
|
|
| --- |
|
|
| ## 3. State-Relay Protocol (NAN-RELAY/1.0) |
|
|
| ### 3.1 State-Packet Structure |
|
|
| ```json |
| { |
| "packet_id": "sp_8f3a92c1d", |
| "mission_id": "msn_001", |
| "sequence": 2, |
| "from_agent": "nan://global/research/web-researcher/2.1", |
| "to_agent": "nan://kw/legal/contract-analyst/1.5", |
| "context": { |
| "Research Agent": { |
| "output": "...", |
| "timestamp": "2026-03-01T09:14:22Z" |
| }, |
| "_summary": "Rolling mission summary..." |
| }, |
| "partial_output": "Agent A output here", |
| "confidence": 0.91, |
| "grounding": { |
| "result": "passed", |
| "sources_verified": ["https://source1.com", "https://source2.com"], |
| "confidence_delta": 0.02 |
| }, |
| "is_rerouted": false, |
| "expires_at": "2026-03-01T09:19:22Z", |
| "mesh_signature": "NAN-SIG:sha256:7f2a...b4c9", |
| "issued_by": "nan://mesh/orchestrator/core/1.0" |
| } |
| ``` |
|
|
| ### 3.2 Relay Flow |
|
|
| ``` |
| Agent A Output |
| │ |
| ▼ |
| ┌─────────────────────────────┐ |
| │ Grounding Gate │ ← verifies claims via MCP web search |
| │ (hallucination check) │ |
| └─────────────────────────────┘ |
| │ PASS │ FAIL → Rerouter evaluates failover |
| ▼ |
| ┌─────────────────────────────┐ |
| │ State-Packet Factory │ ← merges A's output into context |
| │ create_packet(...) │ |
| └─────────────────────────────┘ |
| │ |
| ▼ |
| ┌─────────────────────────────┐ |
| │ Orchestrator Router │ ← selects next agent from pipeline |
| │ relay_engine.relay(...) │ |
| └─────────────────────────────┘ |
| │ |
| ▼ |
| Agent B receives full State-Packet |
| (no context loss, no re-onboarding) |
| ``` |
|
|
| --- |
|
|
| ## 4. Grounding Gate (Grounding-Gate/1.0) |
|
|
| ### 4.1 Trigger Conditions |
|
|
| The Grounding Gate activates on every agent output that contains: |
| - Numerical claims (years, percentages, financial figures) |
| - Recency signals (`"current"`, `"latest"`, `"as of"`) |
| - Attribution phrases (`"according to"`, `"reported by"`) |
| - Legal/regulatory references |
|
|
| ### 4.2 Verification Method |
|
|
| Claims are verified via **MCP web search** (2026 standard): |
|
|
| ```python |
| # MCP tool call |
| { |
| "type": "web_search_20260101", |
| "name": "web_search" |
| } |
| ``` |
|
|
| ### 4.3 Pass/Fail Thresholds |
|
|
| | Condition | Action | |
| |---|---| |
| | ≤ 2 claims fail verification | PASS | |
| | > 2 claims fail verification | FAIL → trigger Rerouter | |
| | No verifiable claims found | AUTO-PASS | |
| | `require_grounding=False` | SKIP | |
|
|
| --- |
|
|
| ## 5. Dynamic Rerouting |
|
|
| ### 5.1 Reroute Triggers |
|
|
| | Trigger | Severity | Action | |
| |---|---|---| |
| | Agent unreachable | HIGH | Immediate reroute | |
| | Grounding Gate FAIL | CRITICAL | Immediate reroute | |
| | Confidence < 0.70 | MEDIUM | Reroute if alternative available | |
| | Latency > 2× SLA | LOW | Log warning, continue | |
|
|
| ### 5.2 Agent Scoring (for reroute selection) |
|
|
| ``` |
| score = (trust_tier_weight × 0.40) + |
| (trust_score × 0.30) + |
| (domain_match × 0.20) + |
| (latency_score × 0.10) |
| ``` |
|
|
| ### 5.3 Trust Score Updates |
|
|
| After every mission step, the Rerouter adjusts the agent's trust score: |
| - `+0.02` on success |
| - `+0.01` bonus for meeting latency SLA |
| - `−0.05` on failure |
| - `−0.03` additional penalty for severe latency breach (> 2× SLA) |
|
|
| --- |
|
|
| ## 6. Authentication (NAN-HS/1.0) |
|
|
| All agent-to-mesh communication uses: |
| - **mTLS** — mutual TLS with 15-minute rotating certificates |
| - **JWT** — signed with Ed25519, 15-minute expiry |
| - **Packet Signature** — SHA-256 HMAC on packet payload |
|
|
| Production environments additionally require: |
| - ADGM Digital Asset registry anchoring for T1_SOVEREIGN agents |
| - PDPL-compliant data residency enforcement for Kuwait jurisdiction |
| |
| --- |
| |
| ## 7. Compatibility |
| |
| | Standard | Version | Notes | |
| |---|---|---| |
| | LangGraph | ≥ 0.2.0 | State machine orchestration | |
| | PydanticAI | ≥ 0.0.14 | Schema validation | |
| | MCP | ≥ 1.0.0 | Web search + tool calls | |
| | FastAPI | ≥ 0.115.0 | Orchestrator API | |
| | Python | ≥ 3.11 | Type system requirements | |
| |
| --- |
| |
| *AgenThink · Kuwait City · ADGM-Registered · MIT License* |
| |
