This repository showcases projects generated using an experimental agentic software engineering framework designed for long-horizon autonomous development. The framework orchestrates LLMs through structured cognitive pipelines—including task classification, strategy selection, iterative critique–refine loops, anti-pattern detection, cross-session learning, and autonomous code execution with quality-gated completion—allowing it to iteratively move from specification to architecture, code synthesis, validation, and spec–code reconciliation while maintaining coherent multi-file systems across hundreds of reasoning iterations. The example projects in this repository are not intended primarily as research contributions themselves, but as demonstrations of the framework’s ability to autonomously prototype complex systems spanning multiple AI paradigms; for example, a quantum neuro-symbolic machine learning framework integrating knowledge graphs, differentiable logic, concept bottleneck models, and hybrid quantum–classical methods. Agentic AI that already works on a weird, quantum neuro‑symbolic fringe—so on standard ML and software problems, it should only perform stronger. Development followed a human-in-the-loop workflow using Claude Sonnet for implementation and Claude Opus for architectural review. To the author’s knowledge, Devin is the closest comparable production system, though this framework takes a different architectural approach to cognitive orchestration and persistent learning. The framework itself is proprietary and is not included in this release.
Initial prompt:
Explore these:
Neuro-Symbolic AI: Combining neural networks with symbolic reasoning; differentiable logic programming; knowledge-guided neural architectures; concept bottleneck models
Quantum Machine Learning (QML): Hybrid quantum-classical models; quantum kernel methods; variational quantum circuits; quantum advantage claims
Then explore if this can be done: Quantum neuro-symbolic AI — quantum circuits implementing differentiable logic on KG structures
The entire project was produced in 7 sessions, ~150 minutes of compute time, across 250 agent iterations.