MARA Research atelier GitHub
LangGraph multi-agent research

Research reports from a single query

Four specialist agents gather sources, synthesize findings, verify citations, and compose a finished report. Built as a reading room: airy, cited, and meant to be read slowly.

Open the atelier Architecture
The atelier

Compose a brief

Agents ready

The desk is clear

Pick a topic or write a query. The sequence then moves through research, analysis, and verification before the report is set on the page.

The method

One sequence, four specialists

Each agent keeps a single job. Breadth, synthesis, verification, and form stay separate so the brief can be both wide and careful.

I

Research

Pulls from web search, Wikipedia, arXiv, and related sources in parallel, then lays the raw material on the desk.

II

Analysis

Synthesizes findings into structured insights and recurring themes, without dressing speculation as fact.

III

Verification

Cross-checks claims against source material before they are allowed to stand in the report.

IV

Report

Sets the executive summary, findings, citations, and confidence score in a form meant to be read.

Systems

Architecture

LangGraph orchestration with persistent memory. A studio map of how the brief moves from query to cited page.

Input User query Research topic
Router Orchestrator LangGraph · state management
Agent sequence
  1. I Research Tavily, Wikipedia, arXiv
  2. II Summarizer Synthesis and key findings
  3. III Citation verifier Fact-checking and validation
  4. IV Report generator Structured output and formatting
Memory

ChromaDB · DuckDB · Gemini 2.0 via Vertex AI

From the folio

Sample reports

Example briefs from the research pipeline, set as citation cards with room to read.

Folio 01 Sample

Transformer Architecture Advances in 2025

Attention improvements, mixture-of-experts scaling, and emerging post-transformer designs.

31 sources 24 citations 96% confidence

Key findings

  • Linear attention variants cut complexity from O(n^2) to O(n) with limited quality loss
  • MoE models reach GPT-4-level performance at about 3x lower inference cost
  • State-space models stay competitive on long-context workloads
Folio 02 Sample

CRISPR Gene Therapy: Clinical Trial Outcomes

Phase III results for CRISPR treatments in sickle cell disease and beta-thalassemia.

28 sources 22 citations 92% confidence

Key findings

  • Casgevy shows 93.5% efficacy in eliminating vaso-occlusive crises
  • Off-target editing rates fall below 0.1% with newer guide RNA designs
  • In vivo delivery that skips bone marrow transplant is in Phase I
Folio 03 Sample

Global Climate Policy After COP30

Belem summit commitments, NDC progress, and feasibility of a 1.5C pathway.

35 sources 29 citations 89% confidence

Key findings

  • Only 12 of 195 signatory nations are on track for 2030 NDC targets
  • Climate finance commitments rose to $300B/year from a $100B baseline
  • Methane pledges could lower the warming trajectory by 0.2C by 2050