AutoKagglab — Documentation
AutoKagglab is an autonomous Kaggle research harness: a local, offline-first system that improves leaderboard solutions through a bounded, reproducible scientific experimentation loop — not by generating notebooks, but by continuously observing results, generating hypotheses, running experiments, evaluating them against frozen cross-validation, and remembering what it learned.
The durable intellectual property of this project is the research protocol (competition config, experiment schema, immutable evaluation harness, budgets, append-only research log, memory). Agent runtimes (Goose, Pi, …) and editors (Zed via ACP, …) are replaceable scaffolding around that core.
Current status
| Item | State |
|---|---|
Feature 001-competition-baseline |
Implemented — all 40 tasks in specs/001-competition-baseline/tasks.md are done |
| CLI | kaggle-agent init / validate / baseline / status (see 03-harness-usage.md) |
| Scope | Tabular CSV, binary classification + regression, LightGBM/CatBoost, CPU, local, single user, no network |
| Spec artifacts | specs/001-competition-baseline/ (spec, plan, research, data-model, contracts, quickstart, tasks, checklist) |
| Constitution | .specify/memory/constitution.md — 13 non-negotiable principles |
| Roadmap | 001 done → 002 strategy engine → 003 ensembling → 004 cross-competition memory → 005 external research → 006 multi-agent (see 06-roadmap.md) |
Document map
| Doc | What it covers |
|---|---|
| 01-overview.md | Vision, core philosophy, non-negotiable principles, the research model (Hypothesis → Experiment → Result → Insight) |
| 02-architecture.md | Layered architecture (methodology / runtime / interface / execution), harness internals, workspace layout, domain model, JSON contracts, safety & governance |
| 03-harness-usage.md | What the harness can do today and how to use it: install, CLI reference, config schema, artifacts, guarantees, tests & CI gates |
| 04-automlab-lessons.md | The automlab/ autoresearch-style reference implementation and how each of its mechanisms maps onto this harness |
| 05-development-workflow.md | How we build the harness itself: the SDD (Spec Kit) loop, dev-time vs runtime loops, phase gates, CI, conventions |
| 06-roadmap.md | Future features with rationale and suggestions: strategy engine, ensembling, cross-competition memory, external research, multi-agent — plus resources |
Source material
These docs synthesize the project’s research inputs:
WORKFLOW.md— the SDD workflow, feature roadmap, anatomy of feature 001, CI gates, runtime interaction model, conventionsopenai-research.md— the architecture conversation: vision, layered design (Autoresearch methodology / Goose runtime / Zed cockpit), MVP spec, build-vs-reuse decisions, post-MVP prioritiesTODO.md— the post-MVP roadmap and rationale (strategy engine before forum mining, ensembling, cross-competition memory, multi-agent last)automlab/— a working autoresearch-style reference repo (read-only) used as the behavioral blueprint
Citation
BibTeX citation:
@online{untitled,
author = {},
url = {https://guillaumegilles.github.io/autokagglab/},
langid = {en}
}
For attribution, please cite this work as: