Stop Blaming the Model: Your Agent Harness Is the Real Problem

The core insight. Alex Zhang argues that most agent harnesses—Claude Code, Codex, Cursor—are functionally identical. The real opportunity lies in designing opinionated harnesses that change how models generalize, not in incremental tooling improvements. The recursive language model (RLM) is one such design: a harness where the only tool is code, and every sub-agent call runs inside an IPython kernel with context offloaded to a file system. This lets the model decompose tasks into locally in-distribution subtasks, even when the full problem is out of distribution.

Why this matters now. Zhang points out that frontier models are already capable enough to handle most tasks reliably for a month, but current harnesses fail to deliver that. He calls this a "skill issue"—not a model capability problem. The evidence: Harvey, a legal AI company, post-trained an RLM on legal document work and got strong results. The RLM abstraction also powers Prime Agent, which uses a persistent sub-agent system and a "continual harness" that lets the agent modify its own tools and system prompt.

The generalization argument. Zhang's key claim is that RLMs naturally generalize across tasks because the meta-strategy (spawn sub-agents, verify solutions) stays the same even when the domain changes. Training on short tasks transfers to tasks 8–30 times longer, and training on math tasks transfers to writing tasks. This is not magic—it's the result of the harness forcing the model to learn a shared solution structure rather than task-specific trajectories.

What's missing. The evidence is still largely anecdotal or based on internal results. Zhang acknowledges that most published results are from collaborations with Prime or Harvey, and that training RLMs at scale requires compute he doesn't have as a PhD student. He also notes that OpenAI's swarm approach for solving Navier-Stokes cost an estimated $40 million, which raises questions about the efficiency of current multi-agent designs.

The trade-off. RLMs are slower and more expensive than compaction-based harnesses for simple tasks. Zhang recommends using simpler approaches for the 95% of cases where they suffice. The benefit of RLMs comes when you need reliable long-running task execution or generalization across very different domains. The cost premium may not be justified for most teams yet.

What teams should watch. Zhang believes we will see dedicated models trained specifically for RLM-like workflows, and that the boundary between model and harness will blur. Teams should track whether Prime or others release post-trained RLM models that are both capable and cost-efficient. For now, the practical value is mostly in understanding the design principles: offload context to code, let the model write its own tools, and treat harness design as a first-class optimization problem.