How it works
What 15 agents find
when they have 2+ days to look.
Not an LLM with a retrieval plugin. The research method is codified as a pipeline, and language models work at its leaves, each call scoped to one sub-question, source, or check.
01 · Intake
Your question is the starting point, not the answer
You describe the question in plain language. We turn it into a structured research plan: clarifying scope, identifying depth, and surfacing implicit assumptions before any research begins.
Under the hood
- Classifies your question by type: open-ended exploration, deep report, literature review, or head-to-head comparison
- Asks follow-up questions to lock down scope and depth before research begins
- Surfaces assumptions you didn't know you were making
- Builds a structured research plan with dependencies between sub-questions
02 · Retrieval
Six channels read at once
Six channels run in parallel, each built to catch evidence the others miss: dense and sparse search over indexed papers, hypothetical-document embeddings, knowledge graph traversal, full-page reading of figures and tables, and the open web. Results are fused, then reranked in two stages.
Under the hood
- Academic papers from PubMed, Semantic Scholar, arXiv, and OpenAlex
- Entities and relationships extracted across papers feed a growing knowledge graph
- Open web crawled through multiple search APIs simultaneously
- Two-stage reranking: relevance filtering, then precision scoring with source metadata
- Detects gaps in coverage and generates follow-up queries automatically
03 · Reasoning
One question, many trees
Your question isn't answered in a single pass. It's split into research topics, and each topic gets its own research tree, up to 4 levels deep with 3 branches per node. Each leaf is investigated by an agent built for that kind of source. Branches are pruned or expanded as evidence surfaces.
Under the hood
- Research tree splits up to 4 levels deep, 3 branches per node
- 15 specialized agents handle query planning, literature research, evidence extraction, adversarial challenge, and more
- Branches that stop producing useful evidence are pruned; promising ones expand
- Findings flow between branches through the knowledge graph
- Different question types route to their own research pipelines
04 · Verification
Three independent signals, one honest score
Every finding gets a confidence score derived from three signals: the agent's own self-reflection, consistency across multiple independent answers, and formal provability via constraint solving. When these signals disagree, the report tells you.
Under the hood
- Each agent scores its own confidence through structured self-reflection
- Independent answers to the same question are compared for consistency
- Quantitative claims are restated as formal constraints and run through an SMT solver, so arithmetic that doesn't hold gets flagged
- Adversarial agents stress-test conclusions and flag contradictions
- Sources from different time periods are cross-checked for temporal consistency
05 · Assembly
Eight phases from findings to finished report
The system plans a narrative arc, writes with inline citations, calibrates tone for the audience, and reviews its own structure for gaps. Each paragraph keeps a provenance record that links it to the source passages it draws on.
Under the hood
- Gathers background context and identifies the narrative structure
- Plans an arc that reads like analysis, not a list of findings
- Validates the outline covers every dimension with no overlaps
- Writes each section with inline citations linked to scored sources
- Calibrates tone and depth for the intended audience
- Reviews its own structure for logical gaps, then renders and delivers the PDF
See it work on your question.
Send us an email with what you want to understand. No account needed.