Thariq Shihipar’s illustration: a viral question asked “why can’t LLMs say which Pokémon end in AW?” — of a thousand Pokémon, two names end in AW (Croconaw and Drednaw). A normal chat model can’t answer, “which is kind of confusing because it definitely knows all the names of the Pokémon.” But Claude Code can, because “it fetches every Pokémon and writes a script to filter for AW.” Anthropic calls this gap capability overhang: “Claude gets smarter in spiky ways” — the knowledge was already there, but reasoning through a thousand names by pure recall-and-filter isn’t reliable, while writing a filter script is. Give it the tool, and dormant capability becomes usable capability.
The same pattern shows up in Claude Code’s own origin story: chat models “had to be given context” and the naive fix was a bigger context window, but “if you give it arms” — the bash tool, the ability to search its own environment — “it can build and search its own context” instead. And Claude Tag’s unlock wasn’t more knowledge either, it was the ability “to work proactively in multiplayer.” In each case the ceiling wasn’t what the model knew, it was what the model could act on. This is the same claim as A mediocre agent inside a strong harness outperforms a stronger agent inside a messy one approached from the model-capability side rather than the product-engineering side: the harness doesn’t just make an agent more reliable, it can be the entire difference between a model that “knows” something and one that can demonstrate it. It’s also why Build for the model six months from now, not the model of today is a coherent strategy rather than a guess — the overhang is already latent in the current model, waiting on the right tool to surface it, not on a future training run. Reducing your unknowns is the trainable skill of agentic coding — and the agent itself is how you train it applies the same unlock-not-teach logic to the human side of the loop: using the agent to surface your own blind spots trains a skill (calibrated self-awareness of your gaps) faster than reading documentation would, the same way giving the model a tool surfaces capability faster than more training would.