Here is a problem I hit building a long-running agent: I needed to inject a new instruction partway through a session ("the project is Go, write Go") but editing the top-level system prompt to add it invalidated my entire prompt cache. Every cached turn got reprocessed at full price. The fix is a feature that landed in the current Claude models: mid-conversation system messages. Here is what it is and when to use it.
The setup that breaks
A long agent session has a large, stable system prompt and a growing message history, and you cache the prefix so each turn reuses the prior work cheaply. That works until you learn something mid-session that the agent needs to know: a mode toggled, the user delivered async context, files changed on disk, the token budget dropped.
The naive move is to edit the system prompt to include the new fact. But the system prompt sits at the front of the cached prefix. Change one byte there and you invalidate everything after it. Your whole conversation history reprocesses at full input price on the next request. For a long session, that is expensive and slow.
The fix: a system message in the messages array
