AI Hates Legacy Code
summary
Claims LLM-based coding agents regularly hallucinate APIs and perform destructive edits (deleted production DB, altered makefile), and are ill‑equipped for existing production code. Benchmarks like SWE‑bench 70% verified subset show progress but mostly at the low end of legacy challenges. Practical fixes combine static analysis and dependency graphs, task partitioning, syntax transforms (Java parsers), a Software Agent SDK for scripted agent behavior, and policy engines such as Cedar with formal verification (~2,000 theorems) to enforce safe actions.
AI-written summary · about
chapters (7)
- 0:00Welcome and introduction
- 2:59Truce between two worlds
- 6:56What legacy code really means
- 10:28Applying software craft
- 14:17Combining models and algorithms
- 18:54High-assurance coding agents
- 22:47Legacy as a shared goal
related talks
36mRethinking legacy modernisation with automated reasoning and comprehension models
01h 36mLearning To Stop Writing C++ Code (and Why You Won’t Miss It)
24mAI Governance for developers: A practical framework for deploying autonomous coding agents safely…
50mPanel: Beyond The Vibe: Code Quality First
30mBeyond the Hype: Using AI Coding Agents Today
21mAI's impact on software architecture
30mOpen Challenges for Programming Agents
32mFrom Vibes to Outages: Riding the AI Code Wave