john@jmchilton.net %
cat problems/agentic-development-legacy.mdx
Thinking about since 2024-11
How do we make agentic development on large legacy systems effective and maintainable?
Applying AI-assisted development to codebases with decades of history, complex dependencies, and institutional knowledge.
ailegacydeveloper-experiencemaintainability
The Question
How do we make agentic development on large legacy systems effective and maintainable?
Why It Matters
AI coding assistants are transforming software development, but most demos show greenfield projects. Real-world impact comes from improving existing systems - the ones with millions of lines of code, implicit conventions, and decades of accumulated decisions. Making AI effective here is harder but more valuable.
Current Thinking
Effective agentic development on legacy systems requires:
- Context management - AI needs to understand not just code, but history and intent
- Safe iteration - Changes must be testable and reversible
- Convention detection - Following existing patterns, not introducing new ones
- Incremental adoption - Can’t rewrite everything at once
The goal isn’t AI replacing developers, but AI amplifying developers who already understand the system.
Open Subquestions
- How much context does an AI need to make good decisions in a legacy codebase?
- How do we prevent AI from introducing inconsistencies with existing patterns?
- What’s the right human-AI division of labor for different types of changes?
- How do we build trust in AI-generated changes to critical systems?