If you've ever worked with an LLM-powered agent long enough for a session to grow, you've probably felt it. Instructions that worked perfectly in a fresh conversation get quietly ignored 15 turns in. The agent starts referencing things it shouldn't.
Source: [Dev.to](https://dev.to/coderonfleek/types-of-context-rot-in-harness-engineering-4fpc)