When NETCONF Is Not Enough: Applicability of MCP and A2A for Advanced Network Management Scenarios

3.1. AI Natural-Language Intent

Operators increasingly expect to issue instructions in natural language: “Raise MTU to 9000 for all Beijing core switches” or “Block source 1.2.3.4 for 30 minutes”. NETCONF requires an edit-config XML blob with exact leaf names and namespaces; even experienced engineers make syntax mistakes under time pressure.¶

The root cause is architectural:¶

• XML is attribute-heavy and case-sensitive; forgotten namespaces or mismatched quotes silently fail.¶
• There is no machine-discoverable “function catalogue” — an LLM must rely on static prompt examples which drift as models evolve.¶
• Multi-vendor differences (OpenConfig vs. proprietary YANG) force the LLM to choose branches inside the XML, exploding prompt size.¶

MCP solves these issues with three primitives:¶

• /tools/list — JSON array of callable functions, each carrying human-readable description and JSON-Schema input.¶
• JSON-Schema — strong-typed, no namespaces, direct mapping to primitive types (string, integer, enum, array).¶
• JSON-RPC 2.0 — single-line request/response, easily parsed by LLM and by controller gateways.¶

Example MCP Tool Descriptor (simplified):¶

{
  "name": "batch_set_mtu",
  "description": "Set interface MTU on multiple devices",
  "inputSchema": {
    "type": "object",
    "properties": {
      "device_group": {"type": "string"},
      "mtu": {"type": "integer", "minimum": 576, "maximum": 9216}
    },
    "required": ["device_group", "mtu"]
  }
}

The LLM now produces:¶

{
  "jsonrpc": "2.0",
  "method": "batch_set_mtu",
  "params": {"device_group": "beijing-core", "mtu": 9000},
  "id": 1
}

3.2. Rapid Model Iteration (DevOps Week-Release)

Cloud-era value-added services must be deployed within days, not months. NETCONF's revision cycle (IETF draft → RFC: 6-9 months) and firmware upgrade windows (≤ 1 per year) are incompatible with weekly release trains. The blocking points are:¶

• YANG module must be burned into firmware before the first config leaf is usable;¶
• Controller regression suite recompiles the entire YANG tree even for a single new leaf;¶
• Backward-compatibility review (must not break old devices) stretches internal QA to weeks.¶

MCP breaks the deadlock by treating "intent" as a hot-swappable Tool rather than a permanent YANG node:¶

• Private YANG is compiled to JSON-Schema in the controller (milliseconds);¶
• Tool registers via /tools/register and is immediately callable;¶
• Gray-list rollout (10 % → 30 % → 100 %) and instant rollback (re-register previous Tool) are done without touching device flash.¶

Example: Cloud-Shield DDoS Cleaning Service¶

Private YANG (120 lines)
  +--rw start-cleanse
  | +--rw target-ip      inet:ipv4-address
  | +--rw bandwidth-Mbps uint32
  | +--rw duration-min   uint16

Compiled JSON-Schema and registered in 30 s:¶

{
  "name": "start_cleanse",
  "version": "1.0.0",
  "description": "Start DDoS cleanse for target IP",
  "inputSchema": {
    "type": "object",
    "properties": {
      "target_ip": {"type": "string", "format": "ipv4"},
      "bandwidth_Mbps": {"type": "integer", "minimum": 100, "maximum": 100000},
      "duration_min": {"type": "integer", "minimum": 5, "maximum": 1440}
    },
    "required": ["target_ip", "bandwidth_Mbps"]
  }
}

Thus MCP is mandatory for any management surface that must support weekly or daily release cycles without waiting for firmware or standards body timelines.¶

4.1. Cross-Controller Long-Flow Orchestration

Maintenance windows for core-network upgrades often exceed 30 minutes and span multiple vendor domains. NETCONF provides atomic configuration on a single controller, but lacks:¶

• a cross-vendor task life-cycle,¶
• human-in-the-loop approval gates, and¶
• delivery of large artifacts (firmware, images, diff reports).¶

A2A fills these gaps with three primitives:¶

• Task — state machine (pending → working → completed/failed/cancelled) persisting across agent restarts;¶
• Artifact — hash-signed object store (≤ 2 GB, resumable upload);¶
• Message — multi-round negotiation (JSON or natural language).¶

Real-World Example: Five-City Core MTU Migration¶

Controllers: Huawei NCE (3 cities) + Cisco NSO (2 cities)
Devices: 312 PE routers
Window: 120 min (02:00-04:00)
Artifact: 2 GB firmware image + 40 MB diff-report

Step-wise A2A Flow (time-stamps):¶

• T+0 min: Orchestrator creates Task T100, goal="Raise MTU to 9000 on core links".¶
• T+5 min: Each controller Agent posts Artifact pre-check.csv (link health KPI).¶
• T+10 min: Orchestrator Artifact hash-verified; human approval card sent to WeChat.¶
• T+15 min: Engineer clicks "approve"; Task state → working.¶
• T+20-90 min: Controllers download 2 GB image via Artifact URL; local NETCONF edit-config issued; progress Artifacts streamed every 5 min.¶
• T+95 min: Last Artifact post-upgrade-verification.csv uploaded.¶
• T+100 min: All agents report success; Task → completed. Total human intervention: 1 click.¶

A2A is mandatory for any multi-vendor, multi-hour workflow that demands task persistence, human gates, and multi-gigabyte artifact delivery—scenarios where NETCONF's single-controller, single-RPC paradigm is insufficient.¶

4.2. Multi-Agent Consensus (Agent-to-Agent)

Fault recovery, security mitigation and resource optimisation often require multiple autonomous agents (monitoring, security, controller, human) to reach a common decision. NETCONF's strict client-server model provides no peer-to-peer capability advertisement, multi-round negotiation or voting primitives.¶

A2A introduces three building blocks:¶

• AgentCard — JSON-LD advertisement of skills and endpoint;¶
• Message — multi-round negotiation (JSON or natural language);¶
• Consensus Engine — policy-based scoring, voting, human-in-the-loop.¶

Consensus Flow Example: DDoS Port Shutdown¶

Agents: Monitor, Security, Controller, Human
Decision: shutdown port 10/1 ?
Scoring: threat_level×0.6 + impact×0.4
Threshold: ≥ 5.0 → shutdown

Message Sequence (time-stamps):¶

• T+0 s: Monitor Agent posts threat_score=9.0 via Message.¶
• T+5 s: Security Agent confirms attack signature; score unchanged.¶
• T+10 s: Controller Agent posts impact=300 VPN down; computed score = 9×0.6 + 3×0.4 = 6.6 (> 5.0).¶
• T+12 s: Task state → awaiting-human-approval; WeChat card sent.¶
• T+135 s: Human clicks "approve".¶
• T+140 s: Controller Agent calls NETCONF shutdown; Artifact post-action.log uploaded.¶
• T+180 s: All agents report success; Task → completed.¶

Therefore A2A is mandatory whenever multiple autonomous agents must discover, negotiate, vote and reach a binding decision — scenarios that NETCONF's unidirectional client-server paradigm cannot emulate.¶