Biology's Central Dogma as a map for the Prompt → Loop → Graph → Evolution trajectory of AI agents.
_{A harness skill — load it into opencode, Claude Code, or Codex CLI to ground agent architecture decisions in biological reasoning.}

---

polymer is a harness skill (first), concept framework (second), and research scaffold (third). It argues the central dogma of molecular biology (DNA → RNA → Protein) describes the same information-flow architecture emerging across AI agent evolution — from simple prompts through loops and graphs toward self-evolving digital organisms.

Unlike a library or runtime, a harness skill is a loadable behavioral lens: inject the prompt from SKILL.md into your coding agent's system prompt, and it will reason about task decomposition, failure modes, and architecture through the biological central-dogma hierarchy. Works with opencode, Claude Code, Codex CLI, and any harness that supports system-prompt injection or skill definitions.

🧬 Quickstart — use as a harness skill

Option A: opencode

// opencode.json (in your project root)
{
  "skills": [{
    "name": "polymer",
    "description": "Central-dogma reasoning for agent architecture",
    "location": "path/to/SKILL.md"
  }]
}

Option B: Claude Code / Codex CLI

# Paste this in your CLAUDE.md or .cursorrules:
#
# You have access to the polymer skill. Before any multi-step
# task, run the 6-layer check: DNA (invariant) → RNA (context)
# → Protein (action) → Regulation (gates) → Graph (topology)
# → Evolution (feedback). See SKILL.md for full prompt.

Option C: Direct injection

Copy the entire core prompt block from SKILL.md (lines 16-51) into your agent's system prompt. The 6-layer check runs automatically before task decomposition.

The isomorphic mapping

Biological	AI analog	What it enables
DNA (genotype)	Model weights, system config, tool definitions	Long-term information storage
DNA replication	Model distillation, checkpointing	Fidelity across generations
DNA repair (exonuclease)	Self-consistency, verifier loops	Error correction
Promoter / enhancer	System prompt, priming	Context-dependent expression
Transcription (DNA→RNA)	Inference forward pass, activation	Transient capability expression
RNA splicing	Context window management, RAG	Multiple outputs from one source
Translation (RNA→Protein)	Structured output, tool calls	From representation to action
Protein folding	Output formatting, chain-of-thought	Functional conformation
Protein function	Tool execution, API calls	Real-world effect
Regulatory feedback	Verifier gates, human-in-loop	Homeostasis, safety
Epigenetics	LoRA adapters, in-context learning	Plasticity without genome change
Horizontal gene transfer	Multi-agent communication (A2A, MCP)	Acquired capability across lineages
Evolution by selection	RLHF, benchmark-driven improvement	Generational improvement

The Loops → Graphs phase shift (July 2026)

On July 18, 2026, Peter Steinberger (@steipete) posted six words that declared the loop engineering era over:

"Are we still talking loops or did we shift to graphs yet?"

575K views in hours. The industry had just named "loop engineering" six weeks earlier — and the conversation had already moved. This repo analyzes that shift through the lens of biological regulatory evolution.

Dimension	Loop	Graph
Control primitive	Think → Act → Observe → Repeat	Nodes + Edges + State
Scheduler	Single-ready (|U| ≤ 1)	Multi-ready (|U| ≥ 1)
Dependencies	Implicit in context	Declared as edges
State	Transcript (growing context)	Typed, checkpointed object
Parallelism	Serial by design	Fan-out / fan-in
Recovery	Restart or context surgery	Retry node, skip, fallback
Audit trail	Scrolling a transcript	Per-edge state transitions
Failure mode	Infinite loop, drift	Structural fragility
Design ethos	"Agent figures it out"	"Declare the structure"

The insight: Loops are not replaced — they are demoted. A loop is a single-node graph with an edge back to itself. Loops live inside nodes; the graph connects them. This is the same relationship transcription (a linear cycle) has to gene regulation (a network of cycles).

The Software Periodic Table connection

The Software Periodic Table identifies ~115 recurring atoms of software (Objects, Properties, Actions, Interfaces, Intelligence, Rules). polymer provides the reaction mechanism — how information flows through those atoms in agent systems. The periodic table gives you the parts; the central dogma gives you the assembly line.

Repository structure

.
├── README.md              ← this file
├── SKILL.md               ← drop-in skill for opencode / Claude Code / Codex
├── RESEARCH.md            ← testable hypotheses, experimental protocol
├── media/                 ← SVG diagrams
│   ├── central-dogma-flow.svg
│   ├── evolution-ladder.svg
│   ├── loop-vs-graph.svg
│   └── evolutionary-tiers.svg
├── docs/
│   ├── 01-central-dogma-map.md       ← full isomorphic mapping
│   ├── 02-loops-to-graphs.md         ← the July 2026 phase shift
│   ├── 03-lambda-calculus.md         ← λ_A, LLMbda formal foundations
│   ├── 04-biological-motifs.md       ← GRN → agent pattern library
│   ├── 05-evolutionary-tiers.md      ← T1–T5 ladder
│   ├── 06-periodic-table-bridge.md   ← how periodic table atoms connect
│   └── 07-practical-patterns.md      ← 10 actionable patterns
└── research/
    ├── references.md                 ← 45 annotated references
    └── open-questions.md             ← 22 open research questions

Status

• Concept: Stable — the central dogma mapping, loops→graphs analysis, and T1–T5 framework are complete
• Skill: Draft — SKILL.md is ready for harness integration
• Research: Scaffold — hypotheses are formulated but untested
• Code: None yet — this is a pure concept repo

---

_{Not production software. A living research scaffold. Push pending inspection.}