Sub-agents are specialized AI components that each handle a narrowly scoped task within a larger multi-agent system. Splitting work across agents helps manage context window limits — modern LLMs suffer from cost scaling (O(n²) attention) and quality degradation ('lost in the middle' effect) as context grows. However, fragmentation creates coordination failures: agents duplicate work, act on stale state, and let errors compound. The solution is shared memory with layered retrieval — short-term session memory, long-term vector-search-backed memory, and coordination primitives like pub/sub and Streams. Redis Iris is presented as a unified engine covering all these needs, with integrations for LangChain, LangGraph, and LlamaIndex, avoiding the vendor sprawl of maintaining separate stores for each memory type.
