Web Research Agent

A Claude Code sub-agent for hotel operators. Cost-efficient web research — finding local businesses, suppliers, products, vendors, and competitive intel — and saving the results so you don't re-research the same thing twice.

Built on the WAT framework (Workflows, Agents, Tools): probabilistic AI handles reasoning, deterministic scripts handle execution.

---

What it does

• Finds local businesses (plumbers, electricians, suppliers, contractors) — returns name, phone, email, address, rating
• Researches products — specs, pricing, availability, where to buy locally
• Pulls competitor / market data — pricing, offerings, reviews of comparable properties
• Returns structured results — tables or bullets with source URLs, ratings, and a clear recommendation
• Caches its findings — every research project is saved to per-agent memory so the next run hits the cache instead of re-searching

---

What you need

• Claude Code installed
• Web access (the agent uses Claude Code's built-in WebSearch and WebFetch tools)
• Optional: an Anthropic API key if you're driving the agent outside a billed Claude Code session

---

Quick start

1. Clone the repo into a Claude Code project:

   git clone https://github.com/th1-ai/web-research-agent.git
   cd web-research-agent

2. Copy the env template and fill it in (or skip it — the agent runs on Claude Code defaults if you have no API keys to wire up):

   cp .env.example .env

3. Edit .claude/agents/web-research.md — replace the {{HOTEL_*}} placeholders described in Customising with your property's details.

4. Open in Claude Code from the project root:

   claude

5. Invoke the agent:

   "Find me three vetted plumbers within 30 minutes of the property who handle commercial work."

   Claude Code will pick up the web-research sub-agent automatically because of the front-matter description line.

---

File structure

web-research-agent/
├── README.md                                    # This file
├── .env.example                                 # Optional API keys
├── .gitignore
├── LICENSE
└── .claude/
    ├── agents/
    │   └── web-research.md                      # The sub-agent definition
    └── agent-memory/
        └── web-research/
            ├── MEMORY.md                        # Index of saved research projects
            └── example-feedback-pattern.md      # Example "memory note" — replace with your own

---

Customising

The agent file uses placeholders so it can be adapted to any property anywhere in the world. Replace the following before first run:

Placeholder	What to put	Example
{{HOTEL_NAME}}	Your property name	The Olive Tree Inn
{{HOTEL_LOCATION}}	Town / village	<your town>
{{HOTEL_REGION}}	Region or state	<your region>
{{HOTEL_NEARBY_TOWNS}}	Comma-separated list of nearby towns	<town A>, <town B>, <town C>
{{HOTEL_LANGUAGE}}	Local search language	French, Spanish, English

You can do this with any text editor or with one sed pass:

sed -i '' \
  -e 's|{{HOTEL_NAME}}|The Olive Tree Inn|g' \
  -e 's|{{HOTEL_LOCATION}}|<your town>|g' \
  -e 's|{{HOTEL_REGION}}|<your region>|g' \
  -e 's|{{HOTEL_NEARBY_TOWNS}}|<nearby towns>|g' \
  -e 's|{{HOTEL_LANGUAGE}}|<local language>|g' \
  .claude/agents/web-research.md

---

How memory works

The agent uses Claude Code's per-agent memory at .claude/agent-memory/web-research/. After each research session, it saves:

• Project files (project_<slug>.md) — one file per research topic, with the question, the answer, contact details, and source URLs. The next run reads these first so it doesn't re-search what's already known.
• Feedback notes — patterns that make future research faster or cheaper (e.g. "for local suppliers, one focused query in the local language beats three broad English ones").

MEMORY.md is the index — keep one line per project file, dated, so stale research is obvious. example-feedback-pattern.md shows the recommended format for a feedback note. Delete it once you have your own.

Heads-up on committing memory files: project research often contains real phone numbers, prices, and email addresses for local businesses. The included .gitignore keeps .env out of git, but does not exclude .claude/agent-memory/. If you want to keep memory private, add .claude/agent-memory/ to your local .gitignore or maintain a private fork.

---

Related agents

This agent works well alongside:

• concierge-agent — pulls in research findings when adding new restaurants, beach clubs, or providers to the vetted-contacts list
• marketing-report-agent — uses competitive research to ground recommendations
• general-purpose-agent — for open-ended research questions that need broader reasoning than web-research alone

---

License

MIT — see LICENSE.