depth-n+ RLM

A depth-n+ RLM is an execution setup where, if an agent gets stuck, it can call up another agent — and that agent, if it too gets stuck, can call up yet another one. Delegation doesn't stop at one level; it can chain through several layers.

Think of it like a workplace. A team lead hands off a task to a team member, that member delegates part of it to a junior, and the junior calls in someone else if needed — the delegation chain doesn't stop at a fixed step. The "n+" in the name means this chain length isn't fixed in advance; it can grow to however many layers (n or more) are actually needed.

Normally, a language model piles up the whole conversation history into one long block of text and searches within it for an answer. The depth-n+ RLM approach instead treats that conversation history like a variable a program can manipulate: it pulls out only the piece of the long record it actually needs and hands the rest of the work off to a sub-agent. The reason for building things this way is that it reduces the chance that small failures during execution get wrongly recorded as a failure of the model itself.

The article describes Prime Intellect's agent harness as following "the recursive language model (RLM) abstraction, treating context as a program and running test-time compute." The easy point to miss here is that this isn't about improving the model's own capability — it's about how the execution shell (harness) wrapped around the model is designed. Even with the same underlying model, benchmark scores can swing significantly depending on whether this structure is in place.