Project AGN

Project AGN (Agent) is an agentic framework for academic research.

Rather than build yet another agent runtime, AGN orchestrates role- specialized Claude Code agents as persistent actors — literature reviewer, experimenter, paper writer, reviewer — wired together with SQLite state and JSONL event logs.

Repositories

agn-paper-machine

An autonomous research-paper generation pipeline. A Python orchestration layer drives a team of role-specialized agents through literature review, experiment design, drafting, and self-review, with persistent memory and explicit message budgeting.

Stack: Python, Claude Code as runtime, SQLite, JSONL events.

agn-shader-framework

A multi-agent framework for GPU and shader R&D — planner, coder, analyst, git-ops, LaTeX reporter, dashboard. Currently early-stage scaffolding for agent-team experimentation.

Stack: Python, uv, pytest, ruff, vectorbt.

Roadmap

• AI conference reviewer — coming soon. An automated reviewer agent for academic submissions, sharing the same role/state primitives as the paper machine.

Status

Mature prototype for the paper machine; early scaffolding for the shader framework. Active development across both.