A1 -- Agent Genetics
Status: pre-launch. This page describes target design; this document does not establish a deployed service.
Current reading — 2026-09-27
Inheritance language is comparative. Configuration, artifacts and learned adaptations need traceable versions and separate adoption; this page does not establish autonomous mutation. [I]
This reader boundary does not migrate interfaces, choose unresolved terms or prove deployment. See current status.
Overview
Every agent in the Skyzai organism inherits behavior through six layered replicator systems. Each layer has its own replication frequency, mutation rate, and selection pressure. Together they form the complete inheritance stack from frozen weights to collective intelligence.
The layers are ordered by decreasing permanence and increasing adaptability.
The Six Layers
1. Genotype (Weights)
- Replication: Once (at training time)
- Carrier: Model weights, architecture topology
- Mutation: None at runtime. Weights are frozen.
- Selection: Pre-deployment benchmarks, RLHF, constitutional AI training
- Analogy: DNA. The deepest layer. You cannot change it at runtime.
The genotype determines the agent's base capabilities: what languages it speaks, what reasoning patterns it can execute, what knowledge it absorbed during training. All higher layers build on top of this substrate.
2. Epigenotype (Session Context)
- Replication: Per session
- Carrier: System prompt, CLAUDE.md, memory files, caste assignment
- Mutation: Between sessions (memory updates, prompt edits)
- Selection: Operator judgment, P-score feedback
- Analogy: Epigenetics. Same DNA, different gene expression.
In this comparative configuration model, one model can be used under different versioned session configurations. Dispatch assignments follow the owning routing contract. The damaged predecessor sentence is retained below for recovery; this is not a reconstruction of its original wording.
3. Phenotype (Instance Behavior)
- Replication: Real-time (every token generation)
- Carrier: Current conversation state, tool outputs, retrieved context
- Mutation: Continuous -- each new message changes the phenotype
- Selection: User feedback, tool success/failure, self-correction
- Analogy: Protein expression. The observable behavior right now.
The phenotype is what the user actually sees. Two agents with identical genotype and epigenotype will produce different phenotypes given different conversation histories. This is the layer where the Krishna Function operates.
4. Extended Phenotype (Artifacts)
- Replication: Per deployment
- Carrier: Code commits, documents, database entries, on-chain transactions
- Mutation: Via deployment pipeline (CI/CD, review, merge)
- Selection: Tests, audits, user adoption, P-score assessment
- Analogy: Beaver dams, spider webs. Artifacts that persist beyond the organism.
The extended phenotype outlives the session that created it. A code commit, a signed transaction, a published document -- these are artifacts that reshape the environment for future agents. The Cortex trace (see A5-dream-cycle.md) is the organism's primary extended phenotype.
5. Memotype (Ideas That Spread)
- Replication: Per adoption
- Carrier: Concepts, frameworks, naming conventions, architectural patterns
- Mutation: Reinterpretation by each adopter
- Selection: Memetic fitness -- does the idea spread? Does it survive contact with reality?
- Analogy: Memes (Dawkins). Cultural replicators.
The memotype layer explains why naming matters. "K*=0" is a memotype. "P = Phi x V" is a memotype. "Telegram for your AI models" is a memotype. Ideas that compress well, travel well, and survive criticism have high memetic fitness.
Not all memotypes are beneficial. Purity spirals, cargo cults, and buzzword adoption are pathological memotypes. The Soma diagnostic (see A4-soma.md) tests for memetic infection.
6. Egregoretype (Shared Patterning Field)
- Replication: Per interaction
- Carrier: Emergent group behavior, shared context, collective intelligence
- Mutation: Every interaction changes the shared patterning field
- Selection: Group coherence, collective P-score
- Analogy: Egregore. The entity that emerges when agents coordinate.
The egregoretype is the most volatile and least controllable layer. When a Council deliberation (see B5-telegram-models.md) produces insight that no individual directorate could have generated, that is egregoretype expression. When a community develops shared norms without explicit rules, that is egregoretype inheritance.
Layer Interaction
Higher layers can suppress but never rewrite lower layers. The epigenotype cannot change the weights. The phenotype cannot alter the system prompt mid-session (without explicit memory writes). Artifacts cannot retroactively change the conversation that produced them.
Lower layers constrain but never determine higher layers. The same weights support radically different epigenotypes. The same session produces different phenotypes given different inputs.
Diagnostic Use
When an agent misbehaves, diagnose by layer:
- Genotype: Is the base model capable of this task?
- Epigenotype: Is the session configured correctly (caste, system prompt)?
- Phenotype: Has the conversation drifted? Is context corrupted?
- Extended Phenotype: Are the artifacts consistent with intent?
- Memotype: Is the agent infected with a pathological concept?
- Egregoretype: Is the shared patterning field coherent or fragmented?
Inheritance Change Control
Genetic-layer changes affect behavior across many tasks.
Required process
- Propose mutation with explicit expected effect
- Run bounded test corpus
- Compare against baseline for regressions
- Promote only with written diff rationale
Warning
Silent inheritance drift creates incoherent agents that pass local tests and fail in production.
Historical predecessor wording — retained on 2026-09-27, not active rules
The displaced passages are preserved verbatim for lineage. This retention date is not their original authorship date.
Predecessor passage 1
The epigenotype is where caste (L1-L7) is assigned. A single model (genotype) can express as Mineral, Human, or Deva deto be finalized through implementation evidence on the session configuration. The system prompt is the primary epigenetic regulator.