Arxiv Paper Coder

arxiv-paper-coder

A multi-agent system that turns a natural-language brief into a working codebase. A planner decomposes the task, a coder writes the files, and a reviewer checks the result — all coordinated by an orchestrator that tracks dependencies, memory, and artifacts. It ships with an arXiv pipeline that builds a daily CS-papers website end to end.

[](https://www.python.org/downloads/) [](https://opensource.org/licenses/MIT)

How it works

Three agents and an orchestrator run a Reason–Act–Reflect loop:

• Planner — breaks the brief into tasks and builds a dependency graph (networkx DAG).
• Coder — implements each task, calling tools for file I/O, code execution, web/arXiv access, and templating.
• Reviewer — checks quality, flags problems, and feeds corrections back into the loop.
• Orchestrator — schedules tasks, resolves dependencies, and manages shared memory and generated artifacts.

                         Orchestrator (core engine)
                Task scheduler · Memory · Dependency resolver
                                   │
            ┌──────────────────────┼──────────────────────┐
            ▼                      ▼                      ▼
       Planning agent  ──▶    Coding agent   ──▶    Review agent
       task decomp,           code gen,             quality checks,
       chain-of-thought,      tool calling,         testing,
       dependency DAG         artifact mgmt         feedback
            └──────────────────────┼──────────────────────┘
                                   ▼
                  Tool layer: file I/O · code execution
                  web search · arXiv API · Jinja2 templates · git

The design borrows from a few familiar patterns: ReAct (Yao et al., 2022), chain-of-thought prompting (Wei et al., 2022), and self-reflection.

Install

git clone https: