AI Research Automation

CAG Deep Research

Causal-Adversarial Graph engine for autonomous research. 7 specialized LangGraph agents with Red/Blue team verification and dialectical judgment — inspired by Popperian falsification.

Source

LangGraphMulti-AgentPydanticOllama

Causal-Adversarial Graph

PlanInvestigateInvestigateSynthesize

Planning causal relationships...

Fig. 1 — Causal DAG with iterative edge investigation. Adversary and supporter agents run in parallel per edge.

LangGraph State Machine

7 Agent Nodes

Causal Planner

Creates initial causal DAG from the research query — maps cause/effect relationships to investigate

Edge Selector

Picks the next causal edge to investigate based on uncertainty and importance

Adversarial Researcher

Red team — searches for evidence that contradicts the current hypothesis

Supporter Researcher

Blue team — searches for evidence that supports the current hypothesis

Dialectical Judge

Resolves conflicts between adversary and supporter evidence, assigns confidence scores

Writer

Synthesizes all validated findings into a coherent research report with citations

Auditor

Safety checks — detects hallucination, prevents loops, validates citation integrity

Methodology

What Makes It Different

▸Causal-Adversarial Graph methodology — models cause/effect relationships, not just information retrieval

▸Red/Blue team parallel verification — adversary and supporter run concurrently, judge reconciles

▸Iterative edge investigation — loops back when fact-checking reveals gaps (not linear pipeline)

▸Hexagonal architecture — agents depend on ports (LLMPort, SearchPort), implementations injected

▸Hybrid LLM routing — round-robin across 11+ Groq models, GitHub Models, Ollama local fallback

▸Action hashing + visit counting — prevents infinite loops and duplicate investigations

Infrastructure

Hybrid LLM Strategy

GitHub Models

GPT-4o, GPT-4o-mini (free with Copilot)

Groq

Llama 3.3 70B, Qwen3 32B, Kimi K2 + 8 more

Ollama

Local Llama 3, Mistral — ~70% cost reduction

Search Providers

TavilyExaDuckDuckGoWikipedia

jayhemnani9910/revolu-idea

7 Agents11+ LLM Models4 Search APIs