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.

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.

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

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

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.