Sub-agents: splitting context across specialized AI agents

What this post added

This post introduces the concept of sub-agents as specialized components within a larger AI system, designed to handle narrowly scoped tasks. It discusses various coordination patterns (orchestrator-worker, supervisor, swarm, router) and highlights the 'context window problem' as a primary driver for splitting agents. The post explains how specialized sub-agents keep context windows focused by sifting through information and returning condensed summaries. It also addresses the challenges of inter-agent misalignment and the need for shared memory to maintain a common ground truth, preventing issues like context poisoning, confusion, clash, and rot. Redis Iris is presented as a solution for providing both memory and retrieval capabilities to these sub-agents, utilizing layered memory (short-term, long-term, episodic, semantic) and hybrid search with vector embeddings and metadata.