Sub-Agent Architectures: Patterns, Trade-offs, and a Kotlin Implementation

Single-agent LLM systems suffer from context bloat, tool confusion, lack of isolation, and no parallelism. Sub-agents address these by using separate LLM calls with isolated conversation contexts and restricted tool sets, though they add latency and cost. The article traces the evolution of sub-agent architectures from ReAct-era tool use and describes AIgent's Kotlin implementation as one design point in a broader space of shared vs. isolated context approaches.