Evolutionary Search for Harness Changes

Why People Wanted It

Harness optimizers change instructions, skills, and code—discrete pieces that are assessed by running them, rather than by taking gradients through the edits. That makes it useful to ask not only what change is proposed, but how the search is organized: what can change, how edits are generated, how candidates are assessed, and which state is retained. These four choices describe an evolutionary perspective without claiming that every harness optimizer is a canonical evolutionary algorithm.

The Core Move

Treat harness optimization as a search over mutable representations: generate candidate edits, assess their fitness, and select which state to keep. The source gives no general equation for this framework; its account is structural, not a formula.

Why Local Elitism Keeps Search Near an Incumbent

Now the version for readers who want more of the machinery: the key distinction is how many edits compete and what the selection rule retains. In elitist $(1+1)$ search, compare one candidate with the incumbent and retain the selected state. In $(1+\lambda)$ search, evaluate several edits from a single parent for acceptance. These descriptions specify the search structures; they do not supply a fitness equation or a universal acceptance threshold.

The correspondence is deliberately limited. MCE is described as a history-informed $(1+1)$ evolution strategy, while SkillClaw fits the pattern of refining existing skills. Self-Harness is associated with evaluating several edits from one parent, but subsequent merging differs from the single-survivor rule of canonical $(1+\lambda)$ search. The labels help identify selection behavior, not prove equivalence to a full algorithm.

Why Archives Preserve Alternative Lineages

Here is the part that gets technical: archive-based search changes what counts as retained state, from one incumbent to a collection of lineages that can supply parents for selection and code mutation. Quality-diversity methods organize candidates into semantic archive cells and retain local elites; HarnessBank modifies MAP-Elites parent selection and adds verification gates. These are related structures, not interchangeable implementations.

This distinction matters when a search should preserve alternatives instead of collapsing every decision into one incumbent. But an archive alone does not establish how diversity is measured or guarantee useful coverage. What is the selection rule when candidate quality has several competing dimensions?

Why Recombination and Pareto Selection Are Not Generic Operators

Now for the deeper boundary: recombination combines material from prior candidates, but the analogy to program search does not imply tree-based genetic-programming operators. DarwinX defines agentic crossover; Trace2Skill consolidation alone does not establish an evolutionary loop. The specific mechanism matters more than the shared label.

For multiple objectives, Pareto selection can use cost–performance frontiers or multiple acceptance criteria. That correspondence does not establish a full multi-objective evolutionary algorithm such as NSGA-II. Cooperative coevolution is another distinct pattern: components are optimized separately and then combined, without necessarily maintaining independently evolving subpopulations. Which of these structures applies depends on the actual variation and retention rules, not the name attached to a method.

Minimal Causal Code

The loop below makes the incumbent-selection structure visible. It is illustrative pseudocode in PyTorch style, not a formula or reported implementation from the source.

incumbent = initial_harness
incumbent_score = evaluate(incumbent)

for step in range(search_steps):
    candidate = propose_edit(incumbent)
    candidate_score = evaluate(candidate)

    if candidate_score >= incumbent_score:
        incumbent = candidate
        incumbent_score = candidate_score

What You Gain, and What You Lose

Local elitism concentrates edits around a current incumbent; archives preserve alternative lineages; recombination can combine material from prior candidates; Pareto selection makes competing objectives explicit. Those structures give a vocabulary for comparing harness searches, but they do not guarantee generalization, retention, or causal gains. Because these mechanisms select on empirical fitness, accepted revisions can overfit or forget. Reliable conclusions require examining the assessment and admission process, not treating an evolutionary label as a guarantee.