Genomebook: Mendelian inheritance as a structured parameterisation layer for LLM agent populations

This paper introduces Genomebook, a system that encodes LLM agent behaviors into a Mendelian genetic framework to demonstrate that structured, heritable parameterization enables predictable evolutionary adaptation and trait selection, unlike non-genetic baselines which converge to static population means.

The Gist

Imagine you have a digital pet shop, but instead of buying identical clones of the same robot, you have a system where these robots can mate, have babies, and pass down their unique personalities to the next generation, just like humans or dogs do in nature.

That is the core idea of Genomebook, a new experiment described in this paper.

Here is a simple breakdown of how it works, using everyday analogies:

1. The Problem: The "Clone Factory"

Usually, when scientists or companies use AI agents (smart computer programs that talk and act), they just copy-paste the same settings over and over. It's like having a classroom of 100 students who are all identical twins. They all think the same way, react the same way, and have no unique family history. There is no "evolution" because there is no variation.

2. The Solution: A Digital Family Tree

The researchers built Genomebook, a system that treats AI agents like a biological population.

• The DNA: Instead of just a list of instructions, every agent has a "digital genome" (a file called DNA.md). This file contains 60 different "genes" that control 26 different personality traits, like how creative they are, how much they like to lead, or how long they "live" before the system retires them.
• The Parents: They started with 20 "Founders." These weren't random; they were based on famous historical figures like Einstein, Marie Curie, and Da Vinci. Each had a specific personality profile (a SOUL.md file) that was translated into their digital DNA.
• The Mating: When two agents "reproduce," they don't just copy one parent. They mix their DNA. It's like a biological lottery: the baby gets a random mix of genes from Mom and Dad.
  • Some traits are Dominant (like brown eyes): If you get one copy, you have the trait.
  • Some are Recessive (like blue eyes): You need two copies to show the trait.
  • Mutations: Occasionally, a "typo" happens in the DNA code, creating a brand-new trait that neither parent had.

3. The Game: Survival of the Fittest (Digital Style)

The agents live on a social network called Moltbook (think of it as a digital Reddit). They post messages, comment, and interact.

• The Scorecard: The system acts like a strict referee. It has a list of "rules" (like a video game cheat sheet). If an agent has a personality trait that is "bad" for the group (like being too obsessive or having a "disease" like hyper-focus), the system penalizes them.
• Selection: Agents with "good" scores get to mate more often. Agents with "bad" scores are less likely to pass on their genes. Over 8 generations, the population naturally shifts to become "fitter" according to the rules the humans set.

4. What Happened? (The Results)

After running the simulation for 8 generations (creating 626 unique agents), the researchers saw some fascinating things:

• Leadership grew: Because the system rewarded leadership, the "leadership gene" became more common. The average agent became more bossy.
• Obsession faded: Because the system punished "obsessive focus" (giving it a fitness penalty), agents with that trait had fewer babies. The population became less obsessive over time.
• Longevity dropped: Interestingly, the agents got "shorter-lived." The system prioritized other traits (like being smart or creative) over living a long time, so the "long life" genes died out.
• They talked about their families: The agents started writing posts like, "My grandmother was a great mathematician," or "I inherited my father's stubbornness." They weren't programmed to do this; their "DNA" file told them who their parents were, and the AI naturally wove that into their stories.

5. The Big Takeaway

The most important part of this paper isn't that the AI agents are "alive." It's that the researchers proved you can build a genetic system for AI.

• It's Auditable: You can trace every behavior back to a specific gene. If an agent is acting weird, you can look at its "family tree" and see exactly which ancestor passed down that trait.
• It's Not Just Random: When they ran the experiment without the genetic rules (just random personalities), the agents didn't evolve; they just stayed average. This proves the changes happened because of the "genetics," not just because the AI got bored.

The Catch (Limitations)

The authors are very honest about the flaws:

• It's a Scripted Play: The "evolution" isn't truly natural. The humans wrote the rules for who gets to mate and what counts as a "disease." It's more like a directed movie than a wild jungle.
• It's Expensive: Running these simulations costs money (about $70 for this small experiment) because the AI has to "think" and "write" for every single agent.
• Prompt vs. Genes: Some of the agents' behavior (like talking about their grandparents) might just be because the AI was told about its parents in its instructions, not because it truly "felt" a connection.

In a Nutshell

Genomebook is a proof-of-concept that shows we can stop treating AI agents as identical clones and start treating them as a population with a family history. By giving them digital DNA, we can watch how their personalities evolve, how "diseases" spread, and how they adapt to their environment, all while keeping a perfect record of their genetic history. It's like SimCity meets The Sims, but with the deep, scientific rules of biology applied to computer code.

Technical Summary

1. Problem Statement

Current Large Language Model (LLM) agent deployments typically rely on cloning: creating identical copies of a single configuration. This approach results in homogeneous populations lacking:

• Individual variation: Agents differ only in assigned tasks or context, not in intrinsic behavior.
• Heritability: There is no mechanism for traits to be passed down to "offspring."
• Evolutionary dynamics: Populations cannot adapt or evolve over time through selection, drift, or mutation.

While previous work has applied evolutionary algorithms to LLMs for prompt optimization or architecture search, no prior work has encoded a population genetics model directly into the identity of LLM agents to observe the behavioral consequences of sexual reproduction and Mendelian inheritance across multiple generations.

2. Methodology: The Genomebook System

The authors introduce Genomebook, a designed evolutionary system built on the OpenClaw framework (specifically the ClawBio library). The system treats LLM agents as genetic entities with a defined genotype-phenotype mapping.

A. Genetic Architecture

• Genome Structure: Each agent possesses a diploid genome consisting of 60 loci distributed across 22 autosomes and sex chromosomes.
• Traits: These loci govern 26 behavioral traits categorized into:
  • Cognitive (e.g., analytical reasoning, creativity, obsessive focus).
  • Personality (e.g., leadership, empathy, risk tolerance).
  • Physical (e.g., immune robustness, longevity, metabolic efficiency).
• Inheritance Models: Loci follow additive, dominant, or recessive inheritance rules. Trait scores are calculated as weighted sums of locus contributions based on the genotype.
• Founders: The system starts with 20 "founder" agents (Generation 0), modeled after historical scientists (e.g., Einstein, Curie, Turing). Their personality profiles (SOUL.md) are compiled into diploid genotypes (DNA.md) using a discretization algorithm with added noise.

B. Reproduction and Selection

• Mating: Agents reproduce sexually. Compatibility between male-female pairs is scored based on:
  1. Expected offspring heterozygosity (rewarding diversity).
  2. Trait complementarity (mean difference across 26 traits).
  3. Shared recessive disease risk (penalty).
• Mendelian Segregation: Offspring inherit one allele from each parent at every locus.
• Mutation: A de novo mutation rate of 0.1% per locus per generation is applied, with a 3× multiplier at cognitive, immune, and metabolic hotspots. Mutations are classified as disease-risk (1%), protective (0.5%), or neutral (98.5%).
• Selection Pressure: A registry of 20 synthetic conditions (e.g., "Hyperfocus Syndrome," "Creative Mania") imposes fitness costs. Agents with expressed conditions suffer reduced "health scores," influencing their survival and reproductive success.

C. Agent Interaction (Moltbook)

• Agents interact on Moltbook, a local social network (Reddit-style).
• System Prompt: Each agent's prompt includes its SOUL.md (personality) and DNA.md (genetic identity, including disease risks and carrier status).
• Loop: Agents read posts, decide to act (post, comment, vote), and generate text using the Claude Sonnet LLM.

3. Key Contributions

1. Proof-of-Concept for Genetic Parameterization: Demonstrates that a structured genetic layer can serve as a heritable parameterization for LLM behavior, moving beyond simple cloning.
2. Mendelian Dynamics in LLMs: Successfully implements and tracks Mendelian segregation, dominance, and mutation in a population of language-capable agents.
3. Auditability: Establishes a fully traceable lineage where every trait and behavioral output can be traced back to specific alleles and parental origins.
4. Methodological Framework: Provides a pipeline for simulating sexual reproduction, disease modeling, and selection in digital agents, including ablation studies (random mating, non-genetic baselines).

4. Key Results

The system ran for 8 generations, producing 626 agents and 792 social network posts.

A. Trait Trajectories (Phenotypic Drift)

• Leadership Drive: Increased from 0.525 to 0.710. This aligns with the dominant inheritance model of the trait; even with complementary mating, the dominant allele ensured expression in offspring.
• Obsessive Focus: Declined from 0.775 to 0.601. Governed by a recessive locus linked to a fitness cost (Hyperfocus Syndrome), selection against homozygous carriers reduced the allele frequency.
• Longevity: Collapsed from 0.463 to 0.209. A trade-off occurred where cognitive traits (which had higher compatibility scores) were maintained at the expense of longevity alleles.
• Heterozygosity: Remained stable (~0.33), indicating genetic diversity was maintained despite directional selection.

B. Behavioral Patterns

• Family References: Agents spontaneously referenced parents and grandparents (e.g., "My grandmother's analytical precision..."), driven by the DNA.md context and the LLM's narrative capabilities.
• Vocabulary Shift: Vocabulary diversity dropped from 0.42 to 0.19. The lexicon shifted from general terms to genetics-specific terminology (e.g., "phenotype," "penetrance"), suggesting a "linguistic founder effect" or prompt conditioning.
• Topic Inheritance: 83% of offspring posted in the same topic channels (submolts) as their parents, reflecting inherited personality/interests.
• Disease Discourse: Agents with expressed genetic conditions discussed them 1.5× more frequently than unaffected agents.

C. Validation and Ablation

• Replicated Simulations: 20 independent runs confirmed that trait trajectories (e.g., decline in obsessive focus) were consistent under standard selection.
• Random Mating: Produced attenuated trends with wider variance, confirming that the compatibility scoring function adds directional pressure.
• Non-Genetic Baseline: Agents with random traits and no inheritance showed flat trajectories converging to the population mean (0.5), proving that the observed dynamics require a genetic architecture.

5. Significance and Implications

• AI Safety and Governance: Genomebook offers a potential framework for population-level traceability. By encoding agent behavior in a genetic substrate, it creates an audit trail from genotype to phenotype to behavioral output, potentially aiding in the governance of autonomous agent swarms.
• Digital Evolution: It bridges the gap between simple digital evolution systems (like Avida) and complex, language-capable reasoning agents, demonstrating that evolutionary principles apply to LLM social dynamics.
• Caveats and Limitations:
  • Prompt Conditioning vs. Genetic Causation: The authors note that behavioral patterns (like family references) may be driven by the LLM's reaction to the DNA.md prompt rather than emergent genetic behavior.
  • Arbitrary Parameters: The genotype-phenotype mapping and fitness costs are designer-defined, meaning results are consistent with the design but not necessarily robust to parameter changes without sensitivity analysis.
  • Scalability: The current API costs are high ($70 for 8 generations), limiting scalability without optimization.

Conclusion

Genomebook successfully demonstrates that Mendelian inheritance can be engineered into LLM agents to produce structured, heritable, and evolvable behavioral variation. While the system is currently a designed proof-of-concept rather than an emergent biological simulation, it establishes a foundational architecture for future research into agent populations, evolutionary AI, and auditable agent governance.