Sisyphus Academica — The Research Paper Writing Army

We’ve all seen what happens when you ask a standard LLM to "write a research paper." You get a highly polished piece of absolute nothing. It spits out smooth, flavorless prose, synthesizes generic summaries, and confidently hallucinates citations that don't exist.

The problem isn't the underlying model; it's how we're using it. We are treating LLMs like advanced word processors when we should be treating them like a fully staffed research lab.

Frustrated by the current state of "AI writing assistants," I decided to build a system focused on research architecture rather than faster typing. I call it Sisyphus Academica—a self-coordinating swarm of 20+ specialized autonomous agents designed to find the gaps in literature that human researchers often miss due to domain bias.

Here is how it works, and why it doesn't write like a bot.

The Problem with Single-Agent Systems

Most AI writing pipelines rely on a single agent or a simple linear sequence: Outline → Draft → Edit. This lack of friction is exactly why the output feels hollow. Humans don't write good research papers in a vacuum; we write them through a messy process of argument, pushback, revision, and rigorous peer review.

Sisyphus Academica introduces intentional architectural friction through a pipeline-based swarm.

1. The Moat: 6 Novelty Engines

Before a single word of the paper is drafted, the core concepts are aggressively stress-tested by six specialized engines designed to inject true academic novelty. Three of the most active include:

• The Contrarian: Forcefully inverts baseline assumptions and claims to see if a stronger, counter-intuitive thesis holds weight.

• The Cross-Pollinator: Drags in methodologies and frameworks from completely unrelated disciplines to solve domain-specific deadlocks.

• The Heretic: Generates fringe, high-risk hypotheses and scores them ruthlessly against existing empirical literature to find unexplored gaps.

2. The Gatekeepers: 10 Adversarial Reviewers

Once a draft is compiled by our 5-agent writing swarm, it doesn't just get formatted into LaTeX. It gets sent to an Adversarial Review Board composed of 10 distinct persona agents.

Your draft has to survive independent, harsh evaluations from The Skeptic, The Methodologist, The Ethicist, and seven others. If the paper fails to satisfy even one reviewer, the system triggers a feedback loop, sending the draft back to the drawing board with explicit critique logs.

3. Absolute Trust: Zero-Hallucination Citations

To solve the academic AI hallucination crisis, the pipeline features a mandatory two-source verification check for every single reference. If an agent asserts a claim and provides a citation, a verification agent cross-checks it across independent repositories. If it cannot be validated by two distinct sources, the citation is entirely stripped from the paper.

4. Overcoming the "AI Signature"

Reviewers can spot standard LLM phrasing from a mile away (the endless em-dashes, the words "delve," "testament," and "beacon"). Sisyphus Academica bypasses AI-pattern detection by applying 41 specific humanizer patterns directly at the token level. Furthermore, it allows you to inject your own writing samples to calibrate the final voice to match your unique stylistic constraints.

The Blueprint

The architecture relies on a modular, Python-based infrastructure built for heavy orchestration:

• Orchestration Framework: Built using OpenCode and OhMyOpenAgent for complex multi-agent state management.

• Writing Swarm: 5 specialized agents deeply tuned with stylistic and humanizer patterns.

• Output Native: Generates clean, compilation-ready LaTeX.

What's Next?

This isn't a tool designed to churn out low-tier papers at scale. It’s an architecture experiment meant to push the boundaries of machine-driven discovery.

I’ve just open-sourced the entire repository. If you are an agent developer, researcher, or AI enthusiast, I’d love for you to spin up the pipeline and throw a topic at it. I am particularly eager to hear how the Heretic engine handles edge cases in highly specialized domains.

👉 Check out the code and contribute here: Sisyphus Academica on GitHub