Building AI Agents for Domain Research: Architecture That Verifies Before It Recommends

A domain research agent that suggests startup.ai without checking the registry is worse than useless—it creates false confidence.

Production agents in 2026 follow one rule: no recommendation without a registry receipt. Everything else—generation, scoring, portfolio alerts—is orchestration around that constraint.

What a domain research agent actually does

Not a chatbot. A pipeline with tool access:

| Capability | Tool layer | Output | | --- | --- | --- | | Generate candidates | LLM (Claude, GPT, DeepSeek) | Base names + rationale | | Verify availability | RDAP → DNS → WHOIS cascade | Boolean + timestamp | | Enrich | WHOIS/RDAP metadata, pricing APIs | Age, registrar, premium flag | | Score | Rules + optional LLM ranker | 0–100 brandability | | Alert | Cron + registrar API | Renewal / drop notices | | Execute | Registrar cart API (human-gated) | Registration draft |

Reference implementations: Vacant (open-source RDAP cascade), naamkaran-ai (domain + trademark grounding), namemy.app MCP (agent tool calls from IDE).

Six-layer architecture

Layer 1: Intent parsing

Natural language → structured brief:

{
  "product": "AI compliance SaaS",
  "tone": "enterprise",
  "tlds": ["com", "ai"],
  "max_length": 12,
  "ban_patterns": ["-ly", "get", "hub"],
  "language": "en"
}

NewName.ai's /api/ai-suggest uses similar constraint injection—including single-word-only rules when users request「单词域名」.

Layer 2: Generation

LLM emits base names without TLD. Batch size: 8–500 depending on cost cap.

Cost receipt: DeepSeek/OpenAI calls at ~$0.001–0.01 per generation batch vs human analyst at $50+/hour. Generation is cheap; verification is the bottleneck.

Layer 3: Availability cascade (critical)

Vacant documents the production pattern:

1. RDAP via IANA bootstrap (authoritative when available)
2. DNS A/NS lookup (active records = taken)
3. WHOIS fallback for TLDs RDAP misses
4. Optional Domainr or registrar API for edge TLDs

NewName.ai /api/check-domain uses DNS NS → A fallback—fast for bulk UI, but production agents should prefer RDAP first where supported.

Concurrency: Vacant limits parallel lookups with shared cache—100 domains as one server request, not 100 client requests (rate limits + IP bans).

Layer 4: Enrichment

Attach metadata only for available names (don't waste WHOIS quota on taken domains):

• Registrar retail price (varies 2–10× on .ai)
• Premium tier flags
• Historical sales from NameBio API comps
• Trademark risk score (naamkaran uses Gemini + Google Search grounding)

Layer 5: Ranking

Weighted score example:

| Signal | Weight | Source | | --- | --- | --- | | Availability verified | Gate (must pass) | RDAP/DNS | | Length 5–10 chars | 20% | Rule | | Pronounceability | 25% | Phoneme heuristics | | LLM brandability | 25% | Model score | | TLD fit (.ai for AI product) | 15% | Rule | | Price under budget | 15% | Registrar API |

Transparency: Log weights in output so humans can override ("I want longer names").

Layer 6: Human gate + optional execution

Agent outputs ranked shortlist with receipts:

1. complynx.ai — AVAILABLE (RDAP 2026-07-28T12:04Z) — $89/yr — score 87
2. reguard.ai — AVAILABLE — $89/yr — score 82
3. auditly.com — TAKEN (NS: ns1.markmonitor.com)

Registration remains human-click unless you implement registrar OAuth with spend caps.

Example workflow: AI productivity tool

Goal: 10 available .ai or .com names under 10 chars, enterprise tone.

| Step | Action | Time | | --- | --- | --- | | 1 | Parse intent → JSON brief | <1s | | 2 | LLM generate 200 base names | 3–8s | | 3 | Rule filter → 120 candidates | <1s | | 4 | Bulk RDAP/DNS on .ai + .com | 15–45s | | 5 | Enrich 34 available with pricing | 5–10s | | 6 | Rank → top 10 | <1s | | 7 | Human picks 1; agent drafts cart link | manual |

Manual equivalent: 4–6 hours of spreadsheet work.

MCP integration: agents in the IDE

namemy.app exposes MCP so Claude Code/Cursor agents call naming + availability as tools during scaffolding:

User: "Scaffold a landing page for my AI legal tool, pick a domain"
Agent: tool_call → generate_names → check_availability → returns complynx.ai

This is the 2026 pattern: domain research as agent tool, not separate browser tab.

Pitfalls that fail production

| Pitfall | Symptom | Fix | | --- | --- | --- | | LLM-only availability | "codepilot.ai is free!" (it's not) | Never surface unverified names | | Client-side bulk check | IP rate limited/banned | Server-side cascade + cache | | Opaque ranking | User distrusts suggestions | Publish weights + raw signals | | Auto-register | Agent buys typo domain | Human approval + spend cap | | WHOIS privacy leak | Search intent exposed to registry | Query via broker API; batch server-side |

Gartner's agent governance guidance applies: read-only tools first, write access (registration, DNS) only after shadow mode.

Privacy and compliance

• Bulk WHOIS queries may violate registrar ToS—use RDAP where possible
• GDPR/PIPL: don't log user search prompts with PII in plaintext
• China: query .cn via domestic registrar APIs; separate from Verisign path

See WHOIS Privacy Guide for registration identity protection.

Build vs buy

| Approach | Cost | Best for | | --- | --- | --- | | NewName.ai + bulk check | Free tier / account limits | Founders naming one product | | Vacant self-hosted | API key + server | Developers wanting full control | | Custom LangGraph agent | $5K–$30K build | Studios naming 10+ products/month | | Enterprise broker workflow | $500+/mo | Portfolio investors 1000+ names |

Most startups should buy until naming volume exceeds 50 searches/week.

China context

• Integrate 阿里云/腾讯云 availability APIs for .cn/.com.cn
• Add 企查查/天眼查 trademark tool call in enrichment layer
• WeChat public account name check as optional Layer 4b
• Domestic LLM (Qwen/DeepSeek) for Chinese single-word constraints

Related workflows

• Generation pipeline: AI-Powered Domain Generation
• Business ops agents: AI Agents for Business Operations
• TLD choice: .ai TLD Boom

Bottom line

Domain research agents are verified search pipelines, not creative writing tools.

Build (or buy) the cascade first. Add generation second. Add ranking third. Keep registration behind a human button.

The moat is not "we use AI"—it is registry truth at scale.