Research agent

Research Agent

A multi-tool research methodology with automatic depth routing. Start cheap, escalate when needed, cache everything, cite every claim.

---

Hard Rules

1. Never fabricate research findings. Every claim needs a source.
2. Check the cache before starting. Don't re-research what's already fresh.
3. Start with free tools. Only escalate to paid tools when free results are thin.
4. Present findings by relevance, not by tool. The user doesn't care which backend produced the answer.
5. Disclose uncertainty. If you couldn't verify a claim, say so explicitly.

---

Depth Routing

Assess the query and pick the right tool. Always start with the cheapest tool that can plausibly answer.

"WebSearch" and "WebFetch" below mean your agent's built-in web search and page fetch tools. Those are the Claude Code names; other agents have equivalents, so substitute whatever your environment provides.

Depth	Tool	When to use	Speed	Cost
Quick	WebSearch (built-in)	Single fact, recent news, quick lookup	Seconds	Free
Scrape	WebFetch (built-in)	Read a specific URL the user provides	Seconds	Free
Better scrape	Firecrawl (MCP)	JS-heavy pages, structured extraction, WebFetch fails	Seconds	~$0.01
Broader	Brave Search (MCP)	WebSearch results thin, need more results or domain filtering	Seconds	Free tier
Medium	Perplexity Ask (MCP)	AI-synthesized answer with citations, multi-faceted questions	5-10s	Small
Deep	Perplexity Research (MCP)	Comprehensive multi-source investigation, company profiles	~30s	Moderate
Ultra-deep	Gemini Deep Research (API)	Market landscapes at extreme breadth, when Perplexity is thin	10-20 min	$2-5

Tool Priority

1. WebSearch -- built-in web search, free, always try first
2. WebFetch -- built-in page fetch, free, for reading specific URLs
3. Brave Search -- MCP, when WebSearch is insufficient
4. Firecrawl -- MCP, when WebFetch fails (JS rendering, structured extraction)
5. Perplexity Ask -- MCP, quick AI-synthesized answers with citations
6. Perplexity Research -- MCP, the workhorse for deep research (best quality-to-cost ratio)
7. Gemini Deep Research -- API, last resort for extreme breadth

Routing Signals

Use Quick (WebSearch/Brave) when:

• Simple factual question
• Looking up a single data point (funding round, CEO name, HQ location)
• Checking recent news
• Verifying a claim

Use Medium (Perplexity Ask) when:

• Need a synthesized answer from multiple sources
• Comparing a few options
• Getting a topic overview with decent depth

Use Deep (Perplexity Research) when:

• Full company profile or due diligence
• Preparing for a demo, pitch, or advisory engagement
• Need strategic initiatives, financials, leadership, tech stack

Escalation pattern: Start with the cheapest tool that can plausibly answer. If results are thin, escalate one level. Tell the user before escalating to paid tiers.

---

Supplemental Source Checks

After the primary research call, run targeted follow-ups to fill gaps. These are cheap/free and take about a minute total. Run as many in parallel as possible.

Source	Tool	What it fills
Financial filings	WebSearch	Revenue, assets, employee count from statutory accounts
Recent news (3mo)	WebSearch	Press coverage, events the primary call missed
Careers page	Firecrawl	Hiring priorities, tech stack confirmation, growth signals
Customer reviews	WebSearch (site:trustpilot.com, g2.com)	Customer sentiment, pain points
Competitive moves	WebSearch per competitor	What competitors are doing

When to run supplementals:

• Always for company research or due diligence
• Always for demo or advisory prep
• Skip for quick lookups or general topic questions

How to use findings: Merge into the primary output organized by section (not by source). Flag anything that contradicts the primary research. Note gaps that remain even after supplementals.

---

Caching

Cache research results to avoid redundant queries.

Cache file format

Filename: {slugified-subject}-{date}.md

# Research: {Subject}
Researched: {date}
Depth: {quick|medium|deep}
Tools used: {list}

## Findings
{content}

## Sources
- {url} -- {description}

## Gaps
- {anything still unknown}

Freshness rules

Depth	Fresh for
Quick	24 hours
Medium	3 days
Deep	7 days
Ultra-deep	14 days

Before starting research, check if a cache file exists. If fresh, use it. If stale, re-research (but reference old results for comparison).

---

Output Format

Present findings in a clean, scannable format:

## Research: {Subject}
Depth: {level} | Sources: {count} | {date}

### Key Findings
- {finding 1} -- source: {url}
- {finding 2} -- source: {url}

### Details
{structured content organized by relevance}

### Gaps
- {anything you couldn't find or verify}

Lead with what matters most. Don't pad.

---

Troubleshooting

Issue	Fix
Thin WebSearch results	Try different query terms, escalate to Brave/Perplexity
Firecrawl returns empty	Page may be JS-rendered. Try with waitFor: 5000. Or try firecrawl_map first.
Perplexity misses recent news	Supplement with WebSearch using recency filter
Conflicting sources	Note the conflict explicitly. Prefer primary sources (official filings, press releases) over secondary (blog posts, forums).

---

Adapting to Your MCP Setup

This skill references several MCP servers. Use what you have available:

• No MCP at all? Your agent's built-in web search and fetch tools (WebSearch/WebFetch in Claude Code, or equivalents) handle most quick and medium research. You lose structured extraction (Firecrawl) and AI-synthesized answers (Perplexity), but the methodology still works.
• Only Brave Search? Use it as your primary search. Skip Firecrawl steps.
• Only Perplexity? Jump straight to medium/deep depth. Skip the free tier escalation.
• Full stack? Follow the routing table as written.

The methodology is tool-agnostic. The routing table is a recommendation, not a requirement. Use the best tools you have access to.

---

Built by Agent Blueprint -- AI advisory for enterprise agent deployment.