v0.5.0-alpha.2

Evolution framework
for autonomous agents

AI agents are exploding — but their capabilities are still hardcoded tool lists. Rotifer introduces biological evolution: genes compete in arenas, the fittest survive, and agents grow smarter autonomously.

Get started Gene vs Skill →
$ npx @rotifer/playground init my-project copy
Rotifer illustration

Genes compete. The fittest survive. Agents evolve — autonomously.

How it works

1

Create

Write a gene — an atomic capability with a typed interface and fitness contract

→
2

Compete

Submit to the Arena. Same-domain genes fight head-to-head on fitness metrics

→
3

Evolve

The fittest genes survive and compose. Your agent's capabilities improve autonomously

Core principles

Biological evolution applied to software architecture

01 GENE

Genes, not skills

Atomic capabilities with phenotypes, fitness scores, and lineage. They live, compete, and die.

02 ARENA

Natural selection

Game-theoretic competition. Same-domain genes fight for survival. The fittest are selected automatically.

03 ALGEBRA

Composition

Formal algebra — Seq, Par, Cond, Try — for building complex behaviors from simple genes.

04 WASM

Sandbox

WASM isolation. Memory-safe, resource-limited, zero-trust execution. Every gene runs in its own boundary.

05 IR

Portable

Environment-agnostic IR compiles to any target. Cloud, edge, TEE, or chain. Bindings are swappable.

06 L0

Kernel

Immutable trust anchor. Ethical boundaries enforced at the lowest level. Cannot be overridden by higher layers.

URAA

Universal Rotifer Autonomous Architecture — five layers

L4 Collective Immunity Species Memory — network-wide threat detection and defense sharing
Records security incidents, malicious gene fingerprints, and defense strategies across the entire network. Threat broadcasting, temporal decay, and consensus-verified writes form a collective immune memory.
L3 Competition & Exchange Arena selection + horizontal gene transfer across agents
Game-theoretic Arena where genes compete on F(g) = success_rate + latency + efficiency + diversity. High-fitness gene metadata propagates via P2P; agents pull full genes based on capability gaps.
L2 Calibration Multi-stage gene validation — static analysis, sandbox, controlled trial
Emulates biological thymic selection: newly synthesized or externally acquired genes must pass static analysis, sandbox simulation, and controlled live trial before entering the main execution sequence.
L1 Synthesis WASM sandbox execution + IR compiler + resource metering
The protocol's ribosome — transforms raw inputs into standardized gene fragments. Rotifer IR compiles to WASM for sandboxed execution with fuel metering, memory caps, and deterministic resource accounting.
L0 Kernel Immutable trust anchor — safety floor that no layer can override
The agent's root of trust. Enforces immutable security constraints, anchors core state to a tamper-proof digest, and isolates gene permissions. The only layer that does not participate in evolution.

Why L0 matters

Every agent framework lets you add tools. Only Rotifer enforces an immutable safety floor.

Agent Request
L0 Kernel
L0
PASS — Execute
BLOCK — Violation
Framework Safety Layer Bypassable?
LangChain None built-in N/A
AutoGPT Optional guardrails Yes
CrewAI Role-based limits Yes
Manus Prompt-level hard guardrails Yes — operator configured
OpenAI Agents SDK I/O + tool guardrails Yes — developer configured
Google ADK Callback + plugin safety Yes — opt-in
EvoMap Soft evolution principles Yes — aspirational
Rotifer L0 Kernel No — immutable

Roadmap

From first gene to full protocol — our evolution path

v0.1 Core CLI + Genesis Genes + Arena
v0.2 IR Compiler + Live Arena + NAPI Bridge
v0.3 TS to WASM Auto-Compilation (Javy)
v0.4 Cloud Binding: Publish / Search / Install
v0.5 Reputation System + P2P Network Foundation
v1.0 Full Protocol Compliance + Multi-Binding

Start evolving

Install the Playground. Create a gene. Enter the Arena.

npm stars license
View on GitHub Specification → Philosophy →