Evolutionary computation was designed when producing a candidate was work.

An algorithm encoded a population, perturbed its members, evaluated them, selected survivors, and repeated. Mutation had to be engineered carefully because variation was scarce and generated inside the optimization loop.

Generative models invert that economy.

Programs, hypotheses, designs, models, explanations, experiments, and strategies can now be proposed faster than we can understand them. Humans generate candidates. Agents generate candidates. Optimizers generate candidates. Running systems reveal accidental candidates whenever conditions expose an unexpected behavior.

Candidate production is becoming abundant. The surrounding machinery for evaluating, relating, and retaining those candidates has not kept pace.

The old loop

A classical evolutionary algorithm moves a synchronized population through evaluation, selection, mutation, and recombination. Its individuals share a representation. The environment and genotype-to-phenotype map are given. Fitness has a known shape. The mutation operator stands outside the population.

Those assumptions make optimization tractable. Long-running adaptive systems tend to violate them.

Real candidates are heterogeneous. They encounter different environments at different times. Their evaluations have costs and uncertainty. Their representations change. Some are compositions of other candidates. Some remain useful as detectors despite being poor actions. Some should sleep for months and return when the context changes.

The machinery proposing and judging them can learn too.

A better covariance update inside CMA-ES touches one part of this change. The larger question is what evolution becomes when candidates, environments, evaluators, representations, and mutation operators all persist.

The population already exists

Every active project contains a population, although it is rarely treated as one.

There are competing implementations, abandoned branches, rival explanations, parameterizations, notebooks, prompts, tests, models, diagrams, and partially successful experiments. Some are explicit alternatives. Others differ in ways nobody recorded.

An AI-assisted project produces this population especially quickly. One conversation may generate dozens of plausible branches. Most become orphaned as soon as the next answer is selected.

Each candidate should retain what produced it, what it inherited, where it behaved differently, which environments evaluated it, what evidence supports or contradicts it, what resources it received, and what descendants learned from the outcome.

This is why lineage is a memory you can walk. Lineage turns accumulated variation into search history.

Evidence and conditions

Fitness is often presented as a number attached to an individual. In a changing environment, evidence belongs to the candidate together with the time, mode, constraints, and conditions under which it was observed.

That evidence can include objectives, behavior, uncertainty, resource use, failures, and consequences. A utility function may reduce it for one decision, but the reduction should not overwrite its source.

A candidate can be dominated under one lens and indispensable under another. It may lose as an actuator and survive as a sensor. It may be uneconomic at one fee tier and efficient at another. It may fail globally while occupying the only known niche that responds coherently to a rare condition.

Selection allocates resources under a stated context. Survival can mean deployment, more evaluation, becoming a parent, retaining a niche, contributing a reusable component, or remaining dormant with a precise reason for return. Losing active resources does not require erasing the memory.

Across scales

The boundary of a candidate need not stay fixed.

A signal transform can be evaluated. So can a policy that consumes several signals, a controller allocating among policies, or a graph containing all three. At a slower clock, the agent modifying that graph and its method for choosing experiments can also vary.

What appears to be an environment from below may appear to be a component from above.

Living systems offer a recurring structural clue: competent units combine into larger units operating on different spatial and temporal scales. The larger unit perceives coarse patterns, supplies constraints downward, and relies on local competence for the details.

Multi-clock intelligence gives this a computational form. Fast artifacts react continuously. Slower evaluators accumulate evidence. Slower agents alter representations, populations, and research programs. A successful insight can be compiled into a fast signal; repeated fast failure can escalate into a change of language or evaluation.

Once a system can inspect its own search process, the mutation operator can enter the candidate set.

An agent may notice that numeric perturbations explore the wrong surface and introduce a structural mutation. It may discover a missing behavioral descriptor, add an evaluation environment, alter the candidate language, or replace the process choosing parents.

This recursion carries an obvious risk. If candidates can reshape their judge while discarding previous conditions, adaptation collapses into selecting environments where the current branch already looks good.

The anchor is the original observation: what was evaluated, under which conditions, against which source, with which information available, and what followed. Evaluators, descriptors, and lenses may change without upgrading old evidence into an experiment that never occurred.

What becomes scarce

When mutation is cheap, candidate count stops measuring progress.

The scarce resources move to environments that can distinguish candidates, exact evaluations, counterexamples, compute, real-world opportunities, attention, and memory. Choosing which uncertainty to reduce can contribute more than generating a hundred nearby children under weak pressure.

Useful diversity also becomes more specific. It appears as consequential disagreement among candidates that retain some competence. Their differences increase what the system can perceive, as described in a map made of reactions.

The population becomes both a repertoire of possible action and an evolving instrument for constructing state.

Persistent populations

Candidates can arrive asynchronously, inherit through composition as well as mutation, and encounter overlapping environments. Resources move toward some and away from others. Useful fragments can cross lineages. Information acquired during operation can be compiled into descendants.

Biology supplies clues rather than a software specification. The practical requirement is a field in which evidence can change what receives attention, what reproduces, what acts, and what becomes part of the substrate.

Making this work requires variation from generative models, heredity from lineage, pressure from evaluation, conditional persistence from niches, and memory from replay.

The next question is where scarce resources and authority should flow. That is the right to act.