Subagents do not share the parent's conversation history. When a subagent starts, it sees only what you put in its prompt. It has no access to files you've discussed, decisions you've made, or context you've built up.
I find this isolation is a feature, not a limitation. It keeps each agent focused and prevents irrelevant context from consuming its token budget. But it means your prompts must be complete.
If a subagent needs to know the project uses TypeScript, tell it. If it needs a specific file path, include it. If it should avoid touching certain modules, say so explicitly. A subagent that guesses wrong can undo work you've already done.