Maria Gorinova: Feedback Loops That Fix Your Codebase

The podcast discusses the evolution and practical implementation of agentic coding within software factories - integrated systems where code, context, and data work together as a cohesive unit. It emphasizes that these systems are not autonomous "black boxes" but require ongoing maintenance, clear workflows, and especially feedback loops to remain effective. Agents have advanced over the past few years, but early versions were limited by lack of context and transparency, making them unreliable. The discussion highlights that successful agentic development depends on providing agents with proper context - such as configuration files, skills, and repository-specific hooks - and establishing well-defined, repeatable processes.

A central theme is the critical role of feedback loops at multiple levels: inner loops for immediate validation (e.g., during PRs), middle loops for pattern-based improvements over time (e.g., reducing excessive comments), and outer loops that respond to production issues. These loops rely on diverse data sources - both human-generated (like PR reviews) and agent-generated (like logs) - to detect recurring problems and suggest targeted fixes. The podcast stresses the importance of starting with narrow, specific goals to avoid ineffective "god loops," and advocates for treating code maintenance as an evolving process where both humans and agents collaborate. Ultimately, sustainable agentic development requires a mindset shift: viewing software factories as dynamic, context-sensitive systems that demand measurement, iteration, and human oversight to manage drift and ensure long-term code health.

• What if you built a feedback loop to automatically detect and fix repetitive code issues in your solo projects?

  • Move: Implement a middle-loop agent that analyzes your recent PRs or commits weekly, identifies recurring feedback (e.g., missing docstrings, lint violations), and opens targeted issues or draft PRs with fixes. Use GitHub Actions to run an LLM-powered script that parses commit messages and comments for patterns.
  • Why Now?: You're already generating the data (commits, comments, CI logs), and ignoring it means manual toil compounds over time. With agentic tools now accessible, you can automate this without infrastructure overhead.
  • Expected Upside: Reduce technical debt accumulation by 40 - 60% and cut review time by catching repeat issues before they reach PR stage.

• What if you treated your personal codebase as a software factory and optimized it for agent maintainability?

  • Move: Audit one active repository to enforce consistency: standardize file structure, add .agentconfig hooks, document decision rationale in READMEs, and modularize code using the Rule of Three. Then, simulate an agent working on a feature to identify friction points.
  • Why Now?: As agents evolve, their performance depends on code context quality - not just your own coding speed. A well-structured solo project increases agent effectiveness if you later automate parts of development.
  • Expected Upside: Improve agent success rate on first attempt by 50%+ and reduce debugging time when integrating AI-generated code.

• What if you used your own PR comments as training data to build a custom code judge for your style and standards?

  • Move: Extract your past PR comments over the last 3 months, cluster them into common themes (e.g., "add validation," "simplify function"), then create a lightweight LLM judge that reviews new diffs for those patterns using a prompt template. Run it in CI as a non-blocking check.
  • Why Now?: You already have unique, high-signal data that reflects your real-world preferences - data most developers ignore. Leveraging it gives you a personalized, low-cost quality gate.
  • Expected Upside: Automate enforcement of 70% of your recurring style and logic feedback, freeing mental bandwidth for higher-level design decisions.

• Implement a narrowly scoped feedback loop (e.g., "comment diet") that analyzes PR comments or code structure to automatically flag and suggest removal of excessive comments in frequently modified files.
• Treat your software factory as an integrated system by co-locating agent configuration (e.g., Claude MD), skills, and hooks directly in the repository to ensure context is version-controlled and reproducible.
• Set up a scheduled middle feedback loop (e.g., weekly) that analyzes historical PR data or agent logs to detect recurring issues, such as missing documentation or repeated review comments, and generates actionable fix suggestions.
• Use an LLM as a lightweight, cost-effective code judge in CI/PR workflows to enforce specific, rule-based standards (e.g., "every public function must have a docstring") that are hard to capture with linters.
• Start with one concrete, measurable pain point (e.g., low test coverage in hot files) and design a feedback loop around it - define the data source, analysis method, and automated action to avoid vague, ineffective "god loops."