MIT’s Alex Zhang discusses recursive language models, multi-agent harnesses, and Prime Agent’s ARC-AGI-3 result

In a recent interview on Latent Space, MIT PhD candidate Alex Zhang discussed recursive language models (RLMs), a class of systems that allow models to treat their own prompts as manipulable objects for recursion and external control. Zhang, first author of the RLM paper, argued that wrapping increasingly powerful models in primitive systems leaves substantial capability on the table, motivating research into RLMs and agent harnesses.

Zhang highlighted Prime Agent, an RLM-based harness designed for long-running autonomous tasks such as coding. According to the interview, Prime Agent was the first system to reportedly solve ARC-AGI-3 before OpenAI’s Astra, leveraging programmatic tool calling, context-as-a-variable, multi-agent messaging, and a self-modifiable harness state to improve token efficiency and expressiveness.

The conversation also covered OpenAI’s internal agent experiment involving approximately 10,000 agents and roughly 130 billion output tokens, which the company used to probe the limits of large-scale multi-agent systems. Zhang contrasted this approach with alternative multi-agent designs, including Kimi’s swarm architecture, and discussed the trade-offs between open-endedness and convergence in agent swarms.

Zhang emphasized harness design as a path to compositional generalization, describing harnesses as scaffolds that can improve performance across diverse tasks by structuring how models interact with tools, context, and subagents. He also explored why systems like Claude Code, Codex, and Pi share structural similarities despite different interfaces, and how future models may hide entire swarms of agents behind a simple user-facing interface.

Beyond RLMs, Zhang touched on context language models (CLMs), which aim to give language models unrestricted control over their own context, and speculated about a future where models reason in a modality-agnostic "Neuralese" rather than natural language alone. He also discussed capability overhang in frontier models and the role of speculative programmatic tool calling in accelerating research.