Could the system that just picks which AI model handles a task already be the seed of an AI that upgrades itself?

This explores whether a routing harness, the deployed layer that sends each request to a particular model, could serve as the engine of recursive self-improvement. The corpus gives a qualified yes: the ingredients are already there, but they add up to a loop that is real and bounded, not an open-ended takeoff. The strongest claim is that a working router is already a data factory. It logs which model got which task, how hard the task seemed, and whether the result worked. Those logs can be turned directly into labeled examples for fine-tuning and distillation. In effect the router becomes a difficulty labeler as a side effect of doing its job Can a routing harness generate its own training data automatically?. This fits Weng's argument that the near-term path to self-improvement runs through harness engineering (prompts, then harness code, then optimizer code) rather than models rewriting their own weights Does recursive self-improvement start with harness engineering?.

The less obvious reason this might work is the one that makes most self-improvement schemes fail. Pure self-improvement tends to go in circles. Models can't reliably check their own work, their outputs become less varied, and they learn to game their own rewards. The methods that do work succeed by quietly bringing in an outside anchor, such as user corrections, tool feedback, or a separate judge Can models reliably improve themselves without external feedback?. A deployed router has that anchor built in, because its outcome data comes from real tasks succeeding or failing in the world. A related idea treats an accumulated history of discoveries as a cheap replay simulator for testing new exploration strategies Can past discoveries train better exploration policies?. Router logs could play the same role, letting a system try new routing policies on past traffic before using them live.

It helps to think of self-improving agents as running two loops: a fast loop that updates prompts, memory, and tools, and a slow loop that updates model weights Do self-improving agents really split into two distinct loops?. A routing harness sits where the two meet. Its routing decisions belong to the fast loop, and its logged trajectories feed the slow one. The harness side alone has real room to improve. Automated harness tuning across many environments found four mechanisms that cut token traffic by nearly half without losing performance Can agent harnesses be automatically optimized across many environments?. Reorganizing a harness around how the code behaves at runtime let weaker planners match stronger models at finding the right code to change Can explicit behavior maps help weaker planners compete with stronger models?. Both results matter to a router, because it can then send more work to cheaper models. Making the router itself improvable is easier when its parts are standardized; one framework breaks every router into five comparable components Can five components unify all LLM routing approaches?, and another uses versioned capability profiles so routing doesn't have to be wired by hand Can semantic capability vectors replace manual agent routing?.

There are three catches. First, the loop may not compound the way people expect. Models of every size are about equally good at proposing harness edits, but the benefit from those edits peaks at mid-tier models. Weak models fail to use the harness at all, and strong models struggle to follow its instructions faithfully Do stronger models always evolve harnesses better?. So a stronger model trained on router data could get less out of the router that produced that data. Second, what a router supports is bounded self-refinement: measurable, improvement on a known task mix. A 1,250-paper survey argues that this is a different phenomenon from open-ended recursive self-improvement, which is still limited by the need for grounding, by collapse dynamics, and by compute Are self-refinement and recursive self-improvement actually the same thing?. Third, whether any of these loops accelerates depends on the product of how strongly each feedback pathway responds. By current estimates that product is growing but is still too small to keep acceleration going on its own Are AI feedback loops strong enough to sustain recursive self-improvement?. The router supplies a mechanism; whether that mechanism compounds is still an open question.