DEV Community

Dan
Dan

Posted on

Compounding Engineering: Turn Your Repo into a Self-Improving DSPy Agent

#programming #ai #python #opensource

Beyond Single-Shot DSPy: Repo-Scale Reasoning That Compounds

I've been experimenting with DSPy beyond one-off prompt optimization. Traditional DSPy shines at optimizing prompts for isolated tasks, but what if your AI agent could learn from your entire codebase over multiple iterations?

Enter Compounding Engineering: a local-first DSPy agent that turns any Git repo into a persistent learning environment. It runs review → triage → plan → learn cycles, building a knowledge base from your code, issues, and past optimizations. No context window limits—improvements compound across sessions.

🚀 Core Innovation

  • Repo as Memory: Indexes your full codebase (Python, JS, configs) into a local vector store. Agents reason over real project context, not toy examples.
  • Compounding Cycles: Each run reviews changes, triages issues, plans fixes, executes via DSPy programs, and learns—storing successes/failures for next time.
  • DSPy-Native: Leverages DSPy signatures, optimizers (BootstrapFewShot, etc.), and metrics. Plug in your LM (OpenAI, local models via Ollama).
  • Local-First: Runs offline with FAISS/Chroma for storage. No cloud dependencies.

Here's the high-level flow:

git clone your-repo
ce init  # Indexes repo, sets up DSPy LM
ce run   # Full cycle: review → triage → plan → learn
ce optimize my_module.py  # Targeted optimization
Enter fullscreen mode Exit fullscreen mode

🎯 Why This Matters for AI Engineers

  • Long-Horizon Planning: Handles repo-scale tasks like "refactor auth module for security" across files.
  • Self-Improvement: Metrics track progress; failed plans become few-shot examples for retries.
  • Open Source Ready: Built for your workflows—integrates Git, DSPy teleprompters, custom signatures.

🛠️ Quick Start 

pip install dspy-compounding-engineering
git clone https://github.com/Strategic-Automation/dspy-compounding-engineering
cd dspy-compounding-engineering
ce init --lm openai/gpt-5.2  # Or your local LM
ce run
Enter fullscreen mode Exit fullscreen mode

Full docs and examples in the README.

🤝 Get Involved

  • ⭐ Star on GitHub if this sparks ideas!
  • Open issues/PRs for features (Rust optimizer? Multi-repo?).
  • Feedback welcome on agentic workflows, long-context reasoning, or DSPy extensions.

Built by Strategic Automation—automating engineering at scale. Let's compound!

Top comments (0)