The Agent Harness as Extended Phenotype: Multilevel Selection, Governed Transfer, and the Hidden Cost of Improving AI Agents

The effective capability of a language-model agent is not determined by model weights alone. This working paper develops a bounded bridge between agent-harness evolution, extended phenotype theory, and evolutionary game theory, then evaluates that bridge in a deterministic mechanism pilot. Multilevel selection does not improve held-out coverage, but eliminates accepted protected-capability regressions and compliance-failure runs at substantially higher evaluation cost. Evidence boundary. This is a synthetic mechanism pilot, not a reproduction of DarwinX and not an empirical validation of extended phenotype theory or evolutionary game theory in deployed AI systems. The accompanying package contains source, frozen hypotheses, raw results, tests, manifest, and provenance evidence.