Skip to content

Architecture

Raghav Kattel edited this page Jun 1, 2026 · 1 revision

Architecture

Sisyphus Academica is a directed acyclic pipeline of specialized agents. Each phase feeds into the next. No cycles — if a phase fails, the paper goes back to the previous phase.

High-Level Diagram

                      ┌──────────────────────────────┐
                      │  Research Director            │
                      │  (orchestrator)               │
                      └────────┬─────────────────────┘
                               │
          ┌────────────────────┼─────────────────────────┐
          ▼                    ▼                          ▼
    ┌───────────┐    ┌────────────────┐    ┌──────────────────────┐
    │ Literature│    │ 6 Novelty      │    │ Gap Analyzer         │
    │ Scout ×5  │    │ Engines        │    │ + Methodology        │
    └───────────┘    └────────────────┘    └──────────────────────┘
                               │
          ┌────────────────────┼─────────────────────────┐
          ▼                    ▼                          ▼
    ┌─────────────────────────────────────────────────────────────┐
    │  Parallel Writing Swarm (5 agents + 41 Humanizer patterns)   │
    └──────────────────────────┬──────────────────────────────────┘
                               │
    ┌─────────────────────────────────────────────────────────────┐
    │  Verifier: Citations × Statistics × AI-Pattern Detection    │
    └──────────────────────────┬──────────────────────────────────┘
                               │
    ┌─────────────────────────────────────────────────────────────┐
    │  Adversarial Review: 10 reviewer personas (parallel)        │
    └──────────────────────────┬──────────────────────────────────┘
                               │
    ┌─────────────────────────────────────────────────────────────┐
    │  Style Auditor: Humanizer certification, em dash zero check │
    └──────────────────────────┬──────────────────────────────────┘
                               │
    ┌─────────────────────────────────────────────────────────────┐
    │  Formatter: LaTeX template, BibTeX, figures, PDF            │
    └─────────────────────────────────────────────────────────────┘

Component Roles

Component Role
Research Director Conducts the whole pipeline. Delegates to subagents. Does not write.
Literature Scout Searches 500+ papers from 4+ sources (arXiv, Semantic Scholar, CrossRef, OpenAlex)
Gap Analyzer Finds genuine research gaps from the literature corpus
6 Novelty Engines Generate novel hypotheses from 6 completely different cognitive frames
Methodology Designer Selects correct statistical tests, runs power analysis
Data Engineer Writes analysis code, generates publication-ready figures
5 Writers Write paper sections in parallel under 41 Humanizer constraints
Verifier 3-module check: citations, statistics, AI-patterns
10 Reviewers Adversarial review from 10 distinct personas
Style Auditor Final certification: zero AI-isms, zero em dashes
Formatter LaTeX compilation, BibTeX generation, PDF output

File Layout

sisyphus-academica/
├── orchestrator/          # Research Director agent
├── subagents/             # 8 core pipeline agents
├── novelty-engines/       # 6 novelty generation agents
├── reviewers/             # 10 adversarial reviewer personas
├── skills/                # Academic Humanizer skill
├── tools/                 # Python CLI toolchain
├── templates/             # LaTeX venue templates
├── config/                # Agent configuration
├── data/                  # Research memory + voice profiles
├── out/                   # Generated papers and figures
├── tests/                 # Python tool tests
└── examples/              # Pipeline output examples

Agent Communication

Agents communicate through files on disk:

  1. Research Director writes a task specification to out/papers/{session}/
  2. Subagent reads the spec, does its work, writes results
  3. Research Director reads results, passes to next phase

This means agent prompts are self-contained — each .md file includes all instructions, output format, and constraints the agent needs. No shared state.

Clone this wiki locally