Paper Implementer

An autonomous agent that reads, breaks down, and re-implements the most-cited machine-learning papers from paperswithcode.com.

How it works

1. Fetch the current "most cited" papers list via the paperswithcode API (scripts/fetch_most_cited.py).
2. Pick the next paper from the queue (papers/tracker.json) — the agent works through the top 10, one or two papers per week.
3. Read → Re-read → Break down → Code → Write up, producing a folder per paper under papers/<year>-<short-title>/ containing:
   • breakdown.md — section-by-section analysis of the paper
   • notes.md — personal reading notes, key insights, questions
   • implementation/ — a working (or partial) from-scratch code re-implementation
   • writeup.md — the agent's own write-up / explanation of the paper
4. Track progress in papers/tracker.json so the next run picks up where the last one left off.

Folder layout

paper-implementer/
├── AGENTS.md                    # Agent operating instructions (the playbook)
├── README.md
├── scripts/
│   ├── fetch_most_cited.py      # Pulls top papers from paperswithcode API
│   └── requirements.txt
├── templates/
│   ├── breakdown.md             # Template for paper breakdown
│   ├── writeup.md               # Template for the write-up
│   └── implementation.md        # Template for the code companion doc
└── papers/
    ├── tracker.json             # Paper queue + progress state
    └── <year>-<short-title>/    # One folder per paper (created per run)
        ├── breakdown.md
        ├── notes.md
        ├── writeup.md
        └── implementation/

Usage

# See the current most-cited papers
python scripts/fetch_most_cited.py --top 10

# Update the tracker with the latest top-10
python scripts/fetch_most_cited.py --top 10 --update-tracker

# Show what's next in the queue
python scripts/fetch_most_cited.py --next

Then point an AI coding agent (Hermes, Claude Code, etc.) at this repo and tell it to follow AGENTS.md.

License

MIT