Can a routing harness generate its own training data automatically?

The paper starts from the requirement that recursive self-improvement needs "a concrete mechanism through which a system observes its own capabilities and converts that evidence into the next round of learning." Its claim is that a deployed routing harness already is one: "beyond task outputs, agentic interaction leaves execution trajectories together with observable evidence of what a model can and cannot yet do." The discussion lists three reusable signals the harness produces while serving requests: execution trajectories that "ground training in real interaction," routing signals that "characterize capability demand," and recorded outcomes that "reveal where the model still falls short."

The mechanism as stated is a heterogeneous model pool behind intelligent routing. For every turn the system records the capability demand predicted, the service tier selected, and the interaction that followed. Those records become user-turn training examples that keep interleaved reasoning, tool calls, and harness context, and they are admitted only after structural validation, six-dimensional semantic evaluation, and subscene-level labeling. The routing signals then do a second job. They "provide estimates of capability demand," which the paper uses to organize supervised fine-tuning into a three-stage curriculum and to extend "naturally" to routing-guided on-policy distillation, where a teacher supervises student-generated responses "under the same staged progression." The router is therefore both a serving component and a difficulty labeler for the training data it helps produce.

This sits on the other side of the slow/fast split from the AIDE2 loop. The AIDE2 note places recursive self-improvement at the harness layer, where the agent rewrites its own code and the excerpt mentions no weight updates (How does an AI agent improve its own research code?). NeoHorse-1 keeps the harness as an instrument and sends what it observes into fine-tuning and distillation, which is the parametric loop in the survey's taxonomy (Do self-improving agents really split into two distinct loops?). SkillClaw starts from the same raw material, interaction trajectories, but converts them into skills synchronized back to agents (How can agent systems share learned skills across users?). My reading, not the paper's: the two-loop taxonomy describes what the conversion step outputs, not where the evidence comes from, since one harness log could feed either loop.

The excerpt does not establish that the loop is recursive. It gives no results, model sizes, or benchmarks, and the abstract frames the system as built "to explore this path." It does not say whether the trained model returns to the pool and shapes later routing and data, which is what would separate a compounding loop from a one-way data pipeline. It does not say who the teacher in the distillation setup is, or whether the routing policy is itself updated. The introduction also notes that trajectories and executable tasks "have already been used to train agentic models," so what the paper adds is the routing signal as a capability-demand label, not trajectory logging. Read at that strength, the note records a design claim: routing metadata gives a deployed system a cheap, per-turn estimate of demand that can order training data. Whether that yields self-improvement is left open by the excerpt.