Claude Dynamic Workflows Are Not Recursive Language Models, Researcher Argues

Here's why Claude's dynamic workflows of 4.8 are not RLM (Recursive Language Model), if you stick to a precise definition, but something new and equally interesting! It took the RLM authors months after the paper to really clarify what was unique about RLM (it's a bundle of ideas) and narrowing that down still seems to be difficult because definitions not used very precisely -- instead people are trying to assign/deduct credit too broadly... (RLM is a solid idea, just the core message getting lost without rigor.) Per common understanding of RLM, looking at new Claude: 1) programmatic access to context instead of stuffing prompt. ❌ (no more than regular tool use, 1M window instead) 2) persistent language REPL storing data over multiple steps. ❌ (writes orchestration script, then runs it) 3) can recursively call the same LLM, with clear return semantics. ✅ (yes, but indirectly via the script) The last two could be scored ~0.5 each, but either way it's only only 1/3 hits at best: so not a RLM... It's more like an evolution of the sub-agent model where the script manages the execution one-shot in an opaque way. IMHO that's a much healthier model for orchestration than some sub-agent implementations -- where it becomes too easy to micro-manage, then quickly gets clumsy & token heavy. https://claude.com/blog/introducing-dynamic-workflows-in-claude-code https://arxiv.org/abs/2512.24601