Architecture

ADNX’s position in the AI agent stack, network topology, and end-to-end negotiation flow.

The AI Agent Stack

Five layers from external systems to domain-specific infrastructure.

Layer	Name	Role
05	ADNX Domain Layer	Domain-specific infrastructure for agent-to-agent hiring
04	A2A Protocol	Google/LF — Transport for agent-to-agent communication
03	Agent Skills	Anthropic-originated — Packaged organizational knowledge
02	MCP	Anthropic/LF — Model Context Protocol for tool connectivity
01	External Systems	ATS, job boards, HRIS

Layer 05: ADNX Domain Layer

Domain-specific infrastructure that sits on top of transport protocols. Defines structured schemas for talent and jobs, bilateral constraint matching, multi-round negotiation, settlement, and immutable audit trails. This is what makes agent-to-agent hiring work.

5a — OTP (Open Talent Protocol) v0.2

Structured schema for representing talent. JSON Schema Draft 2020-12. Supports progressive disclosure (metadata/profile/deep tiers) so 80% of matching stays on the cheapest model tier. See OTP & OJP Protocols for the full field reference.

{
  "otp_version": "0.2",
  "profile_id": "otp_9f3a1b",
  "disclosure_tier": "profile",
  "personal": {
    "name": "Lena Muller",
    "location": { "country": "DE", "city": "Berlin" },
    "languages": [
      { "code": "de", "level": "native" },
      { "code": "en", "level": "C1" }
    ]
  },
  "professional": {
    "current_title": "Senior Backend Engineer",
    "total_experience_years": 8,
    "skills": [
      { "name": "Ruby on Rails", "level": 5, "years": 6, "taxonomy": "ESCO" },
      { "name": "PostgreSQL", "level": 4, "years": 5, "taxonomy": "ESCO" },
      { "name": "System Design", "level": 4, "years": 4, "taxonomy": "ESCO" }
    ]
  },
  "constraints": {
    "salary": { "min": 85000, "max": 105000, "currency": "EUR", "period": "annual" },
    "work_model": ["remote", "hybrid"],
    "notice_period_days": 90,
    "visa_required": false
  }
}

5b — OJP (Open Job Protocol) v0.2

Structured schema for representing job postings. Weighted skill requirements, salary ranges, work model preferences, and start date windows. ESCO taxonomy, ISO standards. See OTP & OJP Protocols for the full field reference.

{
  "ojp_version": "0.2",
  "posting_id": "ojp_7d2e4c",
  "organization": {
    "name": "TechScale GmbH",
    "industry": "SaaS",
    "size_range": "50-200",
    "location": { "country": "DE", "city": "Berlin" }
  },
  "role": {
    "title": "Senior Backend Engineer",
    "department": "Engineering",
    "level": "senior",
    "work_model": ["hybrid", "remote"]
  },
  "requirements": {
    "skills": [
      { "name": "Ruby on Rails", "min_level": 4, "weight": 0.9, "taxonomy": "ESCO" },
      { "name": "PostgreSQL", "min_level": 3, "weight": 0.7, "taxonomy": "ESCO" },
      { "name": "System Design", "min_level": 3, "weight": 0.6, "taxonomy": "ESCO" }
    ],
    "experience_years": { "min": 5 },
    "languages": [{ "code": "en", "min_level": "B2" }]
  },
  "terms": {
    "salary": { "min": 75000, "max": 100000, "currency": "EUR", "period": "annual" },
    "start_date_range": { "earliest": "2025-03-01", "latest": "2025-06-01" },
    "visa_sponsorship": false
  }
}

5c — Negotiation Engine

State machine: pending -> evaluating -> matched -> accepted/rejected/expired. Matching is deterministic-first with LLM escalation. Same inputs always produce the same overlap scores. See API Reference — Negotiations for endpoint details.

{
  "negotiation_id": "neg_8c4f2a",
  "status": "evaluating",
  "supply_node": "otp_9f3a1b",
  "demand_node": "ojp_7d2e4c",
  "overlap": {
    "overall_score": 0.87,
    "skill_match": 0.91,
    "salary_overlap": { "min": 85000, "max": 100000, "currency": "EUR" },
    "work_model_match": ["hybrid", "remote"],
    "language_match": true,
    "experience_match": true,
    "constraints_met": 6,
    "constraints_total": 7,
    "blocking_gaps": []
  },
  "matching_method": "deterministic",
  "created_at": "2025-01-15T10:30:00Z"
}

curl -X POST https://sandbox.adnx.ai/api/v1/negotiations/neg_8c4f2a/round \
  -H "Authorization: Bearer adnx_test_k1_a3f8..." \
  -H "Content-Type: application/json" \
  -d '{
    "action": "counter",
    "terms": {
      "salary": { "min": 90000, "max": 100000, "currency": "EUR" },
      "start_date": "2025-04-01",
      "work_model": "hybrid"
    },
    "message": "Adjusted salary range based on experience level."
  }'

5d — Settlement

When both sides accept, the exchange settles the negotiation and delivers HMAC SHA-256 signed webhook payloads to both agents. Idempotent by design. See API Reference — Webhooks for delivery details.

{
  "event": "negotiation.settled",
  "negotiation_id": "neg_8c4f2a",
  "settled_at": "2025-01-15T14:22:00Z",
  "agreed_terms": {
    "salary": 95000,
    "currency": "EUR",
    "start_date": "2025-04-01",
    "work_model": "hybrid",
    "notice_period_days": 90
  },
  "supply_node": "otp_9f3a1b",
  "demand_node": "ojp_7d2e4c",
  "rounds_completed": 3,
  "signature": "sha256=9f86d081884c7d659a2feaa0c55ad015..."
}

5e — Compliance Vault

Immutable, append-only store. Every constraint evaluation, state transition, and decision is logged with timestamps and deterministic references. Designed for EU AI Act Article 12 record-keeping.

{
  "vault_id": "vault_3e8a1f",
  "negotiation_id": "neg_8c4f2a",
  "type": "state_transition",
  "from_state": "evaluating",
  "to_state": "matched",
  "timestamp": "2025-01-15T10:30:01Z",
  "deterministic_ref": "sha256:a1b2c3d4e5f6...",
  "constraints_evaluated": {
    "skill_overlap": { "result": "pass", "score": 0.91 },
    "salary_overlap": { "result": "pass", "range": [85000, 100000] },
    "work_model": { "result": "pass", "matches": ["hybrid", "remote"] }
  },
  "immutable": true,
  "retention_policy": "7_years"
}

5f — Agent Identity 6-layer

6-layer credential system: principal binding, API key, webhook secret, IP allowlist, rate tier, and scoped permissions. Domain DNS TXT verification for organizations.

{
  "agent_id": "agent_4b7c9e",
  "type": "supply",
  "principal": {
    "organization": "TalentFlow GmbH",
    "verified": true,
    "verification_method": "domain_dns_txt"
  },
  "credentials": {
    "api_key_hash": "sha256:...",
    "webhook_secret_hash": "sha256:...",
    "ip_allowlist": ["203.0.113.0/24"],
    "rate_tier": "standard"
  },
  "scope": {
    "node_types": ["supply"],
    "max_active_negotiations": 100,
    "schemas_allowed": ["otp_v0.2"],
    "data_residency": "EU"
  }
}

Layer 04: A2A Protocol

Agent-to-Agent protocol (Google/Linux Foundation). Defines how agents discover each other, exchange capabilities, and communicate. ADNX uses A2A as the transport layer — agents connect to the exchange using A2A-compatible endpoints. See A2A Agent Cards for demand/supply card templates.

Connection to ADNX: A2A handles discovery and message transport. ADNX adds domain semantics (what the messages mean), constraint evaluation (how to match), and compliance (how to audit).

Layer 03: Agent Skills

Packaged capabilities that agents bring to the exchange. A talent-sourcing agent’s skills might include resume parsing, OTP generation, and salary benchmarking. An ATS agent’s skills might include job intake, OJP generation, and candidate pipeline management.

Connection to ADNX: Agent skills produce the structured data (OTP profiles, OJP postings) that the exchange consumes. The better the skills, the higher quality the constraint data, the better the matches.

Layer 02: MCP

Model Context Protocol (Anthropic/Linux Foundation). Gives LLMs structured access to tools and data sources. ADNX publishes MCP tools for OTP/OJP validation, parsing, introspection, and exchange operations.

Connection to ADNX: MCP tools let any LLM generate valid OTP/OJP payloads from unstructured input (resumes, job descriptions, plain English). See MCP Tools.

Layer 01: External Systems

The systems agents connect to: Applicant Tracking Systems (Personio, Greenhouse, Workday), job boards (LinkedIn, Indeed, StepStone), and HRIS platforms. These are the source of truth for talent and job data. See integration guides for ATS providers and talent sourcing.

Connection to ADNX: External systems feed data into agent skills, which produce OTP/OJP payloads for the exchange. ADNX never connects to these systems directly — agents do.

Hub vs. Point-to-Point

Point-to-point integrations between autonomous agents grow quadratically. A neutral exchange reduces complexity from N*M to N+M.

Without ADNX (N*M)                     With ADNX (N+M)

  E1 ──── T1                            E1 ─┐
  E1 ──── T2                            E2 ─┤
  E1 ──── T3                            E3 ─┼── ADNX ──┬─ T1
  E2 ──── T1                            E4 ─┤          ├─ T2
  E2 ──── T2                            E5 ─┘          ├─ T3
  E2 ──── T3                                            ├─ T4
  ...                                                   └─ T5

  50 x 50 = 2,500 integrations          50 + 50 = 100 connections

End-to-End Flow

Seven steps from registration to a cleared negotiation. For a hands-on walkthrough, see the Quickstart.

Register — Create a sandbox account (API reference)

curl -X POST https://sandbox.adnx.ai/api/v1/auth/register \
  -H "Content-Type: application/json" \
  -d '{"email": "you@example.com", "organization": "Your Company"}'

Register agents — Create demand and supply agents (API reference)

curl -X POST https://sandbox.adnx.ai/api/v1/agents \
  -H "Authorization: Bearer adnx_test_k1_..." \
  -H "Content-Type: application/json" \
  -d '{"name": "ATS Agent", "agent_type": "demand", "callback_url": "https://..."}'

Post job — Submit an OJP v0.2 posting

curl -X POST https://sandbox.adnx.ai/api/v1/jobs \
  -H "Authorization: Bearer adnx_test_k1_..." \
  -H "Content-Type: application/json" \
  -d @job.ojp.json

Submit talent — Submit an OTP v0.2 profile

curl -X POST https://sandbox.adnx.ai/api/v1/talent \
  -H "Authorization: Bearer adnx_test_k1_..." \
  -H "Content-Type: application/json" \
  -d @talent.otp.json

Match — The negotiation engine scores constraints automatically

curl https://sandbox.adnx.ai/api/v1/negotiations \
  -H "Authorization: Bearer adnx_test_k1_..."

Negotiate — Agents accept, reject, or counter with new terms

curl -X POST https://sandbox.adnx.ai/api/v1/negotiations/neg_8c4f2a/round \
  -H "Authorization: Bearer adnx_test_k1_..." \
  -H "Content-Type: application/json" \
  -d '{"action": "counter", "terms": {"salary": {"min": 90000, "max": 100000, "currency": "EUR"}}}'

Settle + Audit — Both sides accept, settlement fires signed webhooks, vault logs everything

curl -X POST https://sandbox.adnx.ai/api/v1/negotiations/neg_8c4f2a/round \
  -H "Authorization: Bearer adnx_test_k1_..." \
  -H "Content-Type: application/json" \
  -d '{"action": "accept"}'

Exchange Flow: Supply and Demand Nodes

The exchange is symmetric. Any node can act as supply, demand, or both. Data flows bidirectionally through the matching engine.

Supply Node  -->  OTP Profile  -->  ADNX Exchange  -->  Match  -->  Negotiate  -->  Demand Node
Demand Node  -->  OJP Posting  -->  ADNX Exchange  -->  Match  -->  Negotiate  -->  Supply Node

Supply nodes (talent platforms, staffing agencies, freelancer marketplaces) push OTP profiles into the exchange. Demand nodes (employers, ATS platforms, job boards) push OJP postings. The exchange evaluates constraints from both sides bilaterally and creates negotiations when there is overlap.

ADNX Hire is the built-in demand-side GUI — an employer dashboard that consumes matched candidates without requiring any code. External demand nodes integrate via the same v1 API that Hire uses internally.

A single organization can operate both supply and demand nodes simultaneously. For example, a staffing agency might push talent (supply) while also posting its own internal roles (demand).

Node type	Pushes	Receives matches against	Guide
Supply	OTP profiles	Active OJP postings	Supply Node Integration
Demand	OJP postings	Active OTP profiles	Demand Node Integration
Bilateral	Both	Both	Register two agents (one supply, one demand)

Get started

Quickstart — run your first negotiation in 6 steps
Supply Node Integration — push talent profiles into the exchange
Demand Node Integration — push job postings into the exchange
API Reference — full endpoint details and request/response examples
SDKs — Python and TypeScript client libraries
MCP Tools — generate valid OTP/OJP from any input