Keeps its parent,
not its anchor

Does the answer an LLM proposer gives most often zero-shot persist inside a search loop? For an agent runtime (a Claude Code subagent requested as Haiku) on circle packing, an AlphaEvolve showcase task, it mostly does not. Without a parent, the subagent's valid outputs at N = 13 and 31 most often match the score of one grid layout, the anchor construction. For single calls that show a parent and ask for a higher sum of radii, we registered that the anchor would persist. Instead, shown that construction as a parent, most valid outputs move above it; shown a parent scoring the family maximum (the best related-grid score), all 26 valid outputs keep or improve its score. Only a parent far below the anchor value often yields the anchor value.

Inside a minimal archive-and-mutate loop, the registered anchor-match rate is lower in descendants than in parentless seeds in all 19 replicate lineages with enough valid outputs to score. Post hoc, most valid descendants keep or improve their parent's score; few reproduce the anchor construction. Validity differs by serving configuration: through a pinned API with thinking off, no directly emitted packing was valid in 105 calls at seven circle counts, unlike most parentless agent-runtime outputs.