Should you read the code, is RAG dead, and did Skills kill MCP?
TL;DR
In a new episode of the GitHub Podcast, the hosts work through current AI debates. Developers should review AI-generated code until they can explain and own the result, with more scrutiny for an authentication refactor than for a CSS experiment. RAG is still alive and remains useful for grounding answers in documentation and codebase context. Skills and MCP are complementary: MCP standardizes tool connections, while Skills package expertise and best practices.
Nauti's Take
The rule of thumb to merge code only once you can explain it is a practical advantage for teams speeding up with coding agents. The limit is time: with large agent diffs, thorough review becomes the bottleneck itself.
We suggest scaling review depth by risk, with strict checks for auth, payments and data access, and tests plus spot checks for UI code.
Sources