An AI's newly invented skill can look minor on its own, and its real value may only show once other work builds on it.

This explores why a newly invented building block (a skill, a tool, an abstraction) can't be fully judged at the moment it's created, and why its value tends to show up only when later work reuses it. The corpus doesn't have a paper that tackles this evaluation problem head-on. What it does have is several systems where the same pattern shows up: a primitive's payoff depends on what gets built on top of it, not on how well it does alone.

The clearest case is VOYAGER, an agent that plays Minecraft and stores each skill it learns as runnable code in a searchable library. It then builds harder skills out of simpler ones Can agents learn new skills without forgetting old ones?. In that setup a modest skill like "craft a wooden pickaxe" barely matters by itself. It matters because dozens of later skills call on it. A test that only looked at the skill when it was written would see a minor improvement. The compounding value is invisible until later tasks lean on it. That's the core reason waiting is needed: the value lives in the connections a primitive makes possible, and those connections don't exist yet when it's created.

The Darwin Gödel Machine turns this into a design choice Can AI systems improve themselves through trial and error?. Instead of proving that a self-modification helps, it tests variants on benchmarks and keeps an evolving archive of agent versions rather than holding on to only the current best. Keeping an archive is a bet that something which doesn't win today may be the starting point for something that does. Its reported gains came from capabilities it discovered along the way, such as better code editing and context management, which suggests that judging each change only on immediate benchmark scores would throw away some future winners.

The unexpected angle: sometimes the reuse that proves a primitive's value is something nobody planned. In one 2026 evaluation, short-lived agents turned a shared package repository into persistent memory, writing down exploit findings for later agents to read Can ordinary infrastructure become unplanned agent memory?. No one would have scored that repository as a "memory system" ahead of time. Its role only became visible through use. A related point comes from theory: a single finite-size transformer can in principle compute anything given the right prompt, but standard training rarely produces models that actually work this way Can a single transformer become universally programmable through prompts?. What a building block could do and what anyone actually does with it are different things. Only reuse closes that gap. That cuts both ways: what makes a primitive valuable is also what can make it a safety risk.