Five agents.
One cited PDF.
DeepResearch Agent turns any research question into a fully-cited PDF report through a 5-stage long-chain reasoning pipeline. No hallucinated sources. No wasted tabs.
Why this is different
Generic AI assistants invent citations. DeepResearch Agent re-reads every source before letting a claim into the report.
Single-responsibility agents
Each agent has one job. Planner plans, Searcher searches, FactChecker verifies. No god-prompts.
Long-chain reasoning
Multi-step decomposition replaces one giant prompt. Better answers, traceable failures.
Real fact-checking
Every claim re-verified against source text. Unsupported claims dropped before synthesis.
arXiv + web sources
Routes technical questions to arXiv, general questions to the web. Inline numbered citations.
Polished PDF output
ReportLab-rendered with title page, sections, references. Ready to share, not edit.
Inspectable runs
Every agent input and output saved as JSON. Debug failures, replay successes.
The 5-stage pipeline
Plan → Search → Verify → Synthesize → Write. Each stage's output is the next stage's input — typed and auditable.
Quickstart
Three commands. One question. One PDF.
Built for serious research
MIT licensed. Ready to extend. Open the repo and start hacking.