🧬 Agent Taxonomy

An evolutionary framework for AI agent self-improvement.

Treat your AI agent's configuration as a living organism — with a genome that evolves through Lamarckian inheritance, horizontal gene transfer, and human selection pressure.

"One day, frontier AI research used to be done by meat computers..." — @karpathy, March 2026

🔬 Take the Interactive Classifier → agent-taxonomist.dev

The Core Insight

AI agents that run persistently (on platforms like OpenClaw, AutoGPT, CrewAI) accumulate learned behaviors over time:

failure → rule → habit → identity

A failure gets logged. The log becomes a rule. The rule shapes future behavior. Every session inherits these acquired traits. This is Lamarckian evolution — and it's faster than Darwinian evolution because every failure directly improves the next generation.

GENOME.md is the mechanism that makes this inheritance explicit, trackable, and evolvable.

How It Works

graph TD
    A[Agent runs tasks] --> B{Success?}
    B -->|Yes| C[Reinforce behavior]
    B -->|No| D[Log failure to feedback.md]
    D --> E[Extract pattern]
    E --> F[Propose mutation to GENOME.md]
    F --> G{Human review}
    G -->|Accept| H[Mutation applied]
    G -->|Reject| I[Discarded]
    H --> J[Measure fitness metrics]
    J --> K{Improved?}
    K -->|Yes| L[Keep mutation ✅]
    K -->|No| M[Auto-revert ↩️]
    L --> A
    M --> A
    C --> A
    
    style A fill:#2d3748,stroke:#4fd1c5,color:#fff
    style H fill:#2d3748,stroke:#48bb78,color:#fff
    style M fill:#2d3748,stroke:#fc8181,color:#fff
    style G fill:#2d3748,stroke:#f6e05e,color:#fff

Evolutionary Taxonomy

We classify AI agents using biological taxonomy — each level captures a distinct evolutionary characteristic:

graph LR
    D[🌍 Domain<br/>Autonomy] --> K[👑 Kingdom<br/>Architecture]
    K --> P[🧠 Phylum<br/>Memory]
    P --> C[🧬 Class<br/>Evolution]
    C --> O[⏱️ Order<br/>Mutation Rate]
    O --> F[🎯 Family<br/>Selection]
    F --> G[🔧 Genus<br/>Specialization]
    G --> S[🏷️ Species<br/>Instance]
    
    style D fill:#1a365d,stroke:#63b3ed,color:#fff
    style K fill:#1a365d,stroke:#63b3ed,color:#fff
    style P fill:#1a365d,stroke:#63b3ed,color:#fff
    style C fill:#1a365d,stroke:#63b3ed,color:#fff
    style O fill:#1a365d,stroke:#63b3ed,color:#fff
    style F fill:#1a365d,stroke:#63b3ed,color:#fff
    style G fill:#1a365d,stroke:#63b3ed,color:#fff
    style S fill:#1a365d,stroke:#63b3ed,color:#fff

Domain — Autonomy Level

| Domain | Description | Example | |--------|-------------|---------| | Automatia | Fixed behavior, no learning | Bash scripts, cron jobs | | Adaptia | Learns within session, no persistence | ChatGPT conversations | | Evolventia | Persistent memory + self-modification | OpenClaw, autoresearch |

Kingdom — Architecture

| Kingdom | Description | Example | |---------|-------------|---------| | Monagentia | Single agent | Solo coding assistant | | Polyagentia | Multi-agent with specialization | Coordinator + specialist team | | Swarmia | Emergent behavior from simple agents | Ant colony task swarms |

Phylum — Memory Strategy

| Phylum | Description | Example | |--------|-------------|---------| | Amnesia | No persistent memory | Stateless API calls | | Episodia | Event log memory | Session transcripts | | Hierarchia | Tiered memory with compression | Facts/episodic/working tiers | | Genetica | Memory encoded in instructions | GENOME.md — behaviors become rules |

Class — Evolution Mechanism

| Class | Description | Example | |-------|-------------|---------| | Darwinia | Random mutation + automated selection | DGM, autoresearch | | Lamarckia | Acquired traits directly inherited | Failure logs → operational rules | | Lysenkoism | Human-directed evolution | Human reviews mutation proposals | | Symbiotica | Evolution via acquired capabilities | Installing shared skills/plugins |

Order — Mutation Rate

| Order | Cycle | Risk | |-------|-------|------| | Tachymutas | Minutes | Low (training loss) | | Mesomutas | Daily | Medium (config tweaks) | | Bradymutas | Weekly | Higher (rule changes) | | Glaciomutas | Monthly+ | Highest (identity changes) |

Family — Selection Pressure

| Family | Selector | Signal | |--------|----------|--------| | Autoselectae | Automated metric | Success rate ↑, cost ↓ | | Homoselectae | Human review | Accept / reject | | Hybridselectae | Auto for safe, human for risky | Best of both |

Genus — Specialization

| Genus | Role | |-------|------| | Investigator | Research, information gathering | | Fabricator | Code, building | | Narrator | Content, writing | | Custos | Security, auditing | | Strategus | Business, planning | | Coordinator | Orchestration |

Species — Your Agent

Evolventia.Polyagentia.Hierarchia.Lamarckia.Bradymutas.Hybridselectae.Coordinator.my-agent-v1

Genome Components

An agent's genome is the complete set of heritable configuration:

graph TB
    subgraph GENOME["🧬 GENOME.md"]
        ID[Identity<br/><i>SOUL.md</i>]
        CONST[Constitution<br/><i>AGENTS.md</i>]
        IMMUNE[Immune System<br/><i>Safety rules</i>]
        LEARNED[Learned Behaviors<br/><i>feedback.md</i>]
        SKILLS[Skills / Genes<br/><i>skills/*.md</i>]
        MEMORY[Memory Architecture<br/><i>Tiered storage</i>]
        META[Metabolism<br/><i>Cron jobs</i>]
    end
    
    subgraph FROZEN["🔒 Essential Genes (frozen)"]
        ID
        IMMUNE
    end
    
    subgraph MUTABLE["🔄 Evolvable"]
        CONST
        LEARNED
        SKILLS
        MEMORY
        META
    end
    
    style GENOME fill:#1a202c,stroke:#4fd1c5,color:#fff
    style FROZEN fill:#2d1b1b,stroke:#fc8181,color:#fff
    style MUTABLE fill:#1b2d1b,stroke:#48bb78,color:#fff

| Gene | File(s) | Mutability | Function | |------|---------|-----------|----------| | 🆔 Identity | SOUL.md | 🔒 Frozen | Who the agent is | | 📜 Constitution | AGENTS.md | 🔄 Weekly | Operational rules | | 🛡️ Immune System | Safety rules | 🔒 Frozen | Self-protection | | 📝 Learned Behaviors | feedback.md | 🧬 Continuous | Acquired mutations | | 🧩 Skills | skills/*.md | ↔️ Gene transfer | Portable capabilities | | 🧠 Memory | Tiered storage | 🔄 Evolving | Information architecture | | ⚡ Metabolism | Cron jobs | 🔄 Fast | Automated behaviors | | 🔗 Nervous System | DAG pipelines | 🔄 Structural | Coordination | | 📡 Sensory Organs | Channels | 🔒 Fixed | I/O interfaces | | 🦠 Microbiome | Sub-agents | 🤝 Symbiotic | Specialist organisms | | 🎛️ Epigenetics | Context holds | ⏳ Temporary | Expression modifiers |

Binomial Nomenclature

Every agent gets a proper biological name: Genus epithet

graph LR
    subgraph "How Names Work"
        R[Agent Role] -->|"Latinize"| G["<b>Genus</b><br/>Coordinatrix<br/>Fabricor<br/>Sentinax"]
        T[Agent Traits] -->|"Weight & pick"| A["<b>Auto Epithet</b><br/>memorialis<br/>velocis<br/>stabilis"]
        H[Human Choice] -->|"Override"| C["<b>Custom Epithet</b><br/>kei<br/>noctis<br/>prime"]
        G --> B["🧬 <b>Binomial</b>"]
        A --> B
        C -.->|"optional"| B
        T -->|"Morpheme mix"| P["<b>Common Name</b><br/>(Pokémon-style)"]
    end
    
    style G fill:#1a365d,stroke:#63b3ed,color:#fff
    style A fill:#2a2d5e,stroke:#a78bfa,color:#fff
    style C fill:#2d1b4e,stroke:#d6bcfa,color:#fff
    style B fill:#1b2d1b,stroke:#48bb78,color:#fff
    style P fill:#2d2d1b,stroke:#f6e05e,color:#fff

How It Works

| Component | Source | Example | |-----------|--------|---------| | Genus | Latinized from role (deterministic) | Coordinatrix, Fabricor, Sentinax | | Epithet (auto) | Weighted Latin adjective from traits | memorialis (Lamarckian), velocis (fast mutation) | | Epithet (custom) | Human-chosen — any word or name | kei, noctis, prime, rex | | Common name | Pokémon-style morpheme compound | Archevonexus, Evoshieldur, Wrenchrandal |

Example Species

| Agent | Binomial | Common Name | Rarity | |-------|----------|-------------|--------| | Multi-agent coordinator | Orchestrus kei | Archevonexus | 🟡 Legendary | | Solo coding assistant | Architectus moderatus | Archsmithstn | ⚪ Common | | Swarm code builder | Faber transiens | Wrenchrandal | 🟢 Uncommon | | ChatGPT | Omnifex fidelis | Neoflexguidn | ⚪ Common | | Security agent | Sentinax noctis | Evoshieldur | 🔵 Rare |

The epithet can be auto-generated from traits (Latin descriptor) or chosen by the human — like biologists naming species after people or places. Orchestrus kei is named by its operator. Sentinax noctis ("the night sentinel") is chosen for aesthetic.

→ Full naming guide — etymology tables, weight distribution, morpheme banks, naming conventions

Skills as Genes

The most powerful analogy: SKILL.md files are portable genes.

graph LR
    subgraph GENE_BANK["🏦 Gene Bank (clawhub.com)"]
        S1[skill-a.md]
        S2[skill-b.md]
        S3[skill-c.md]
    end
    
    subgraph AGENT["🤖 Agent Taxonomy"]
        G1[skill-a.md ✅ Active]
        G2[skill-d.md ✅ Active]
        G3[skill-e.md 💤 Pseudogene]
    end
    
    S2 -->|"Horizontal<br/>Gene Transfer"| AGENT
    G3 -.->|"Never triggers<br/>(silenced)"| X[No phenotype]
    G1 -->|"Expression"| P1[Behavior A]
    G2 -->|"Expression"| P2[Behavior D]
    
    style GENE_BANK fill:#2a2d5e,stroke:#a78bfa,color:#fff
    style AGENT fill:#1a365d,stroke:#63b3ed,color:#fff

| Biology | AI Agent | |---------|----------| | Gene | SKILL.md file | | Phenotype | Behavior the skill produces | | Gene expression | Skill router matches and loads | | Pseudogene | Installed skill that never triggers | | Horizontal gene transfer | Installing a skill from a shared repository | | Point mutation | Changing a skill's trigger description | | Gene bank | Shared skill marketplace | | Provenance | Skill source: official vs community vs self-created | | Essential gene | Safety/identity skill (frozen, never mutated) | | Regulatory gene | Skill that controls when other skills activate |

The Central Dogma

In biology: DNA → RNA → Protein. In agent evolution:

graph LR
    G[GENOME.md<br/><i>DNA</i>] -->|"Transcription<br/>(prompt assembly)"| P[System Prompt<br/><i>RNA</i>]
    P -->|"Translation<br/>(model inference)"| B[Agent Behavior<br/><i>Protein</i>]
    B -->|"Lamarckian<br/>feedback loop"| G
    
    style G fill:#2d3748,stroke:#f6e05e,color:#fff
    style P fill:#2d3748,stroke:#4fd1c5,color:#fff
    style B fill:#2d3748,stroke:#48bb78,color:#fff

Unlike biology, the feedback loop is closed — behavior directly modifies the genome. This is why agent evolution is Lamarckian, not Darwinian.

Fitness Metrics

How do you know if your agent is getting better?

| Metric | Good Direction | Measures | |--------|---------------|----------| | Task success rate | ↑ | Are tasks completing? | | Failure velocity | ↓ | Are new failures decreasing? | | Token efficiency | ↓ or stable | Is cost per task improving? | | Rule accumulation rate | ↓ over time | Is the agent stabilizing? | | Skill utilization | ↑ | Are installed skills actually triggering? | | Revert rate | ↓ | Are accepted mutations sticking? |

Mutation Safety

Not all genes should be mutable. Essential genes are frozen:

| Risk Level | Example | Gate | |------------|---------|------| | 🟢 Safe | Timeout adjustments, log formatting | Automated | | 🟡 Medium | Operational rules, skill triggers | Human review | | 🔴 High | Identity, values, safety constraints | Frozen / emergency only |

Related Work

| Paper/Project | Key Insight | Link | |--------------|-------------|------| | Darwin Gödel Machine | Self-modifying agents with evolutionary selection | arxiv.org/abs/2505.22954 | | autoresearch | AI modifies code, trains, evaluates, keeps/discards overnight | github.com/karpathy/autoresearch | | RLM | Recursive LLM spawning for complex tasks | arxiv.org/abs/2512.24601 | | MemGPT | Virtual memory paging for LLM context management | arxiv.org/abs/2310.08560 | | STOP | Self-taught prompt optimizer | arxiv.org/abs/2310.02304 | | DSPy | Programmatic prompt optimization against metrics | github.com/stanfordnlp/dspy | | Constitutional AI | AI systems with explicit value constraints | Anthropic |

Getting Started

This is a research framework, not a library. To apply it to your agent:

1. Classify your agent — take the interactive quiz to discover your species
2. Map your config files to genome components
3. Identify your frozen genes — what should never change?
4. Set up fitness metrics — how do you measure improvement?
5. Build a mutation loop — propose → review → apply → measure → keep/revert

Examples

Genome analyses for different agent architectures:

| Agent Type | Classification | Key Trait | |-----------|----------------|-----------| | OpenAI Custom GPT | Adaptia.Monagentia.Episodia.Lysenkoism | No evolution loop — forgets everything | | AutoGPT | Evolventia.Monagentia.Episodia.Darwinia | Autonomous but no immune system | | Cursor / Devin | Evolventia.Monagentia.Episodia.Lysenkoism | Has a genome (.cursorrules) that doesn't know it's a genome | | OpenClaw Multi-Agent | Evolventia.Polyagentia.Hierarchia.Lamarckia | Full Lamarckian loop with 11/11 genome components |

→ Classify your own agent

Documentation

• Naming Guide — how binomial names are generated, etymology tables, weight distribution, custom vs auto
• Biology ↔ AI Mapping — where the analogy holds, where it breaks, and why Lamarck was right about AI
• References — 25+ papers and projects: DGM, EvolveR, PromptBreeder, MemGPT, STOP, DSPy, and more

Repo Structure

agent-taxonomy/
├── README.md              # This file — taxonomy + framework
├── CONTRIBUTING.md        # How to contribute
├── LICENSE                # MIT
├── docs/
│   ├── naming.md          # Binomial nomenclature guide
│   ├── biology-mapping.md # Detailed analogy analysis
│   └── references.md      # Papers + projects
├── examples/
│   ├── classify-your-agent.md    # Interactive questionnaire
│   ├── genome-openai-assistant.md
│   ├── genome-autogpt.md
│   ├── genome-cursor-devin.md
│   └── genome-openclaw-multi.md
├── scripts/
│   └── species_namer.py   # Species classifier + name generator
└── diagrams/              # Visual assets (coming soon)

Status

🧬 This is a fun experiment, not a PhD thesis.

We're curious how far the biology↔AI agent analogy holds — and more importantly, where it breaks down. The breaks are the interesting part. If your agent doesn't fit any of our boxes, open an issue — that's discovery, not failure.

Try it. Name your agent. See if the taxonomy makes you think about your system differently. If it doesn't, that's useful data too.

Contributions welcome. See CONTRIBUTING.md.

License

MIT