AI & ML Research 8 Hours

Last updated: May 16, 2026, 11:46 AM ET

LLM Architectures

A new technical analysis dissects recursive language models — a class of agents that invoke themselves iteratively to handle complex tasks — contrasting them with ReAct, Code Act, self-loops, and subagent architectures. Unlike ReAct's alternating reasoning-observation loops, recursive models maintain a single persistent context, enabling deeper chain-of-thought without external tool calls. Code Act extends this by allowing code execution within each recursion, while self-loops and subagents introduce parallel or hierarchical delegation. The piece argues that recursion reduces prompt overhead and improves task decomposition for multi-step reasoning, though at the cost of increased token usage per invocation.