Most arguments about the next leap in AI capability point at the model itself: more parameters, more compute, more curated training environments. A new analysis from independent researcher Alex Zhang makes the opposite case. The bottleneck is not the neural network, Zhang argues, but the program wrapped around it, the harness that decides what the model sees at each step and how it acts.
That claim matters because compositional generalization, meaning a system’s capacity to combine previously learned skills to handle a task rather than requiring something nearly identical during training, is the thing modern post-training has struggled to buy at any price. Labs have responded by curating more environments and running longer training horizons, an approach Zhang calls a brute-force paradigm. If the underlying network still cannot compose what it already knows, every new domain requires its own expensive round of data collection.
Zhang’s evidence centers on a harness design he calls a Recursive Language Model, or RLM, built to offload long or complex context into a code REPL instead of appending everything to one growing prompt. Two mechanisms do the work. Context offloading routes input-specific information through a symbolic variable, keeping it out of the model’s main context. Programmatic sub-agent calling treats tools and sub-agents as functions whose outputs live in memory rather than in the model’s visible history. Together, these let structurally similar tasks look nearly identical to the underlying language model, even when the surface-level problems look nothing alike.
The result, per Zhang’s experiments training the open-weight model Qwen3-30B-A3B-Instruct-2507, configured as an RLM, is that training on short tasks generalized to held-out tasks 8 to 32 times longer, with roughly ten times the evaluation lift compared with training the base Transformer directly at matched training gains. A standard Transformer trained the conventional way did not generalize nearly as well across that length gap, and training on one task domain barely transferred to a different one.
The skeptical read is built into the piece itself. Zhang singles out Claude Code and Codex, two of the most widely used coding agents, as harnesses that fail his own test. Both keep appending tool outputs and reasoning to a single expanding context window, a pattern Zhang ties directly to the “context rot” that operators already notice in long agent sessions, where output quality degrades as the transcript grows even though nothing in the prompt is technically wrong. If Zhang is correct, that degradation is not a scale problem waiting on a bigger model. It is a design flaw in how today’s most popular coding agents manage their own memory.
That framing lands squarely inside AI Insiders’ running argument that the harness, not the underlying model, is where competitive advantage now accumulates. A frontier lab shipping a marginally better base model does not automatically produce a better agent if the scaffolding around it keeps flooding the context window with stale state. A team running a smaller open-weight model inside a harness built around context offloading and programmatic sub-calls could plausibly outperform a bigger model wrapped in a naive prompt-stuffing loop.
Zhang’s results come from a research blog post, not a peer-reviewed paper, and the RLM experiments run on a single 30-billion-parameter model tested against two benchmarks, MRCRv2 and GraphWalks. That is a narrow evidentiary base for a claim this broad, and Zhang has not published code or weights alongside the post. Even so, the practical implication for builders is concrete: before budgeting for a bigger model, audit whether the current harness forces the model to treat every new task as unfamiliar. Teams evaluating coding agents should test how performance holds up as context length grows well past the range they typically operate in, not just at whatever window the benchmark defaults to.
Analysis by Alex Zhang, published July 20, 2026, on his personal research blog.
