Agentic, multi-step web research with sub-agent spawning, evidence-ledger artifacts, and reproducible reports. Use when asked to research, look things up, compare sources, or gather citations across multiple web pages. Triggers: research, look up, search for, find information, compare, gather sources, web research, /search command.
Use for web research that needs:
Trigger via: /search <query> command or when user asks to research something.
| Mode | When to Use | Tool Budget | Stop Threshold |
|---|---|---|---|
| quick | Fast answers, simple questions | 20 calls | Coverage ≥60%, Confidence ≥0.65 |
| standard | Balanced research (default) | 40 calls | Coverage ≥80%, Confidence ≥0.75 |
| deep | Thorough investigation |
| 60 calls |
| Coverage ≥90%, Confidence ≥0.80 |
| product-parity-hunt | Finding integrated stacks/platforms | 60-80 calls | Parity gates satisfied (see below) |
Automatically switch to product-parity-hunt mode when the user's request includes:
When auto-escalating, inform the user:
"This looks like a product-parity hunt. I'll use the
product-parity-huntmode which requires broader coverage across solution classes and stricter evidence requirements."
This mode is for finding pre-made orchestration layers — deployable systems that already include scheduling, durable execution, persistence, UI, and notifications.
Create these files immediately in ~/research/<slug>/:
parity-matrix.md — Define what "parity" means for this huntopen-questions.md — Minimum 12 concrete questionsquery-families.md — Minimum 8 query families plannedcandidate-inventory.md — Empty scaffold with solution-class quotasNo searches or fetches until Step 0 is complete.
| Phase | Name | What Happens |
|---|---|---|
| 0 | Define | Create artifacts, no tool calls |
| 1 | Solution-Class Sweep | Breadth search with quotas (min 2 cycles) |
| 2 | Integration Hunt | Search for starter kits, templates, docker-compose |
| 3 | Deep Verification | Fetch proof pages for shortlisted candidates |
| 4 | Synthesis | Score parity matrix, produce decision |
Before deep dives, the candidate inventory MUST include at least:
Agent frameworks may be included but must not exceed 25% of early inventory.
For each shortlisted candidate, fetch at least:
No marketing-only claims: If the only evidence is a marketing page, parity score cannot exceed 1.
You may NOT stop until ALL are true:
parity-matrix.md exists with scores + notes for every rowcandidate-inventory.md has ≥3 candidates per platform classquery-families.md has ≥8 families and ≥30 executed queriesevidence-ledger.mdIf gates cannot be satisfied and budget is exhausted:
stop-report.md explaining which gates failedIf the last iteration produced:
Then the next iteration MUST expand scope by:
No more "keep fetching more docs from the same framework."
references/tool-policy.md for tool priority and concurrency rules~/research/<slug>/ and returns executive summaryreferences/scaffold.md| Priority | Tool | Use For |
|---|---|---|
| 1 | mcp__brave-search__brave_web_search | Breadth searches, query expansion |
| 2 | mcp__exa__web_search_exa | High-quality extraction |
| 3 | mcp__exa__web_search_advanced_exa | Filtered searches (domain, date, category) |
| 4 | mcp__firecrawl__firecrawl_scrape | JS-heavy sites, clean markdown |
| 5 | mcp__firecrawl__firecrawl_map | Site discovery, docs portals |
| 6 | WebFetch | Quick static pages (cheap) |
| 7 | Browser tools | Interactive/login only |
Use the Task tool with:
subagent_type: "search-specialist"