Features / Research
Deep web research,
multiplied.
Seven search backends, hybrid fusion, hierarchical fan-out and cross-referencing — an answer assembled from the whole web, not the first page.
Capabilities
How Parallel researches the web
Linkup, Parallel.ai, Exa, Tavily, Serper, Brave, DuckDuckGo — with hybrid and smart modes and RRF ranking to fuse results.
spawn_agent at any depth up to max_depth; batches of children run in parallel, background mode reports results as notifications.
web_fetch, web_crawl and web_feed pull full pages, follow sitemaps and parse RSS at up to 1.1M items/s.
parse_html (531k rows/s) and extract_json turn any document into structured data.
Real Chrome via CDP when the page fights back; analyze_image for screenshots, charts and documents.
code_symbols and repo_map map codebases; PDF extraction reads whitepapers and filings.
Comparison
Parallel vs. manual research
Manual
- Hours per topic, one tab at a time
- Single search engine, first-page bias
- Copy-paste into spreadsheets
- No verification, stale data
- Nothing remembered next week
Parallel
- Minutes per topic, parallel agents
- 7 backends fused with RRF ranking
- Structured findings + exports
- Verified contacts, cross-referenced facts
- Semantic memory carries knowledge forward
Under the hood
Multi-agent fan-out explained
Plan
LLM decomposes the query into sub-tasks; subtasks persist in SQLite.
Fan out
Coordinator spawns sub-agents (JoinSet or separate OS processes).
Collect
Budget-capped summaries flow back; context stays small.
Reflect
Goal Mode: an LLM judge checks the result against the goal and runs gap-filling rounds.
Synthesize
Findings merge into summary.md + findings/, absorbed into memory.
Agent roles
Five roles, one pipeline
Every agent runs a role-specific prompt on top of the same runtime. Roles can be overridden per deployment via [agent.role_models].
Decomposes the query into 2–5 sub-tasks, delegates, and synthesizes the final answer.
Web search + OSINT / lead-gen workflows with source citation.
Cross-references findings, spots patterns and contradictions, assigns confidence HIGH/MED/LOW.
Adversarial fact-checking: VERIFIED / LIKELY / UNVERIFIED / CONTRADICTED.
Structures the report: Summary / Findings / Analysis / Sources / Gaps.
Stable / context / volatile prompt layers maximize prefix-cache hit rates across the fleet.
Use cases
Built for real business scenarios
Market & competitor scans
Track pricing, positioning and hiring across an industry — with diffs between runs.
Due diligence
Company facts, news, filings and key people into one verified dossier.
Lead research
Decision-makers at target companies, enriched and verified before outreach.
Tech scouting
Map repos, extract symbols, summarize documentation with code_symbols + repo_map.