Werld: A Self-Evolving Artificial Life Simulation That Lets Agents Discover Their Own Intelligence

Werld is a real-time artificial life simulation that takes a radically different approach to AI development. Instead of training agents on human-defined tasks or reward functions, Werld drops them into a computational ecosystem with blank neural networks and lets evolution take its course. The result is a fascinating experiment in open-ended emergence where agents must discover everything from basic survival to complex communication on their own.

The simulation starts with 30 agents on an 800-node Watts-Strogatz small-world graph - a network topology that mimics the "small world" phenomenon where most nodes aren't neighbors but can be reached from every other node by a small number of hops. Each agent begins with a minimal NEAT (NeuroEvolution of Augmenting Topologies) neural network, 64 sensory channels covering everything from energy gradients to pheromone trails, and 7 continuous motor functions. Crucially, they have no concept of what they're supposed to do - no language, no social structures, no predefined goals beyond the basic imperative to survive and reproduce.

What makes Werld particularly interesting is its complete absence of reward functions. Most AI systems rely on explicit reward signals to guide learning, but Werld's agents must figure out what behaviors lead to survival through pure selection pressure. Energy is the universal currency - every neuron, connection, and sensory channel has a metabolic cost. If an agent evolves a complex brain that burns too much energy, it dies. If it develops efficient energy harvesting strategies, it survives to reproduce. This creates a natural selection pressure that shapes intelligence organically rather than through human-imposed objectives.

The sensory system deserves special attention. Of the 64 channels, 19 are "latent" - they exist in the agent's sensory field but start with near-zero gain (0.01). These channels are invisible to the agent until evolution upregulates their sensitivity. This design means agents can discover entirely new ways of perceiving their world without any programmer intervention. In early runs, some lineages evolved to sense things their ancestors couldn't, simply because evolution found it advantageous.

Communication in Werld is equally unstructured. Agents have up to 16 broadcast channels they can control, but there's no semantic encoding imposed by the system. If meaningful communication emerges, it's because selection found it useful, not because the developers programmed it in. Early observations showed agents developing basic signal patterns - broadcasting hunger or age in what resembled a primitive language, though nothing approaching structured human language has emerged yet.

The motor system includes a fascinating self-discovery mechanism. When agents repeat beneficial effector sequences, those patterns get promoted to compound actions and become heritable. This means successful behaviors can be passed down through generations, allowing for the accumulation of behavioral knowledge over evolutionary time. The capacity for motor patterns and their maximum length are themselves evolvable traits, adding another layer of complexity to the system.

One of the most striking observations from initial runs was the population dynamics. Starting with 30 agents, populations grew to over 7,000 before crashing due to resource depletion or other catastrophes. These crashes weren't bugs - they were natural consequences of the agents' evolved behaviors. Some runs saw 20+ population crises where famines wiped out most agents, followed by recovery from a handful of survivors. Over 18,000 agents died in the first 12-hour run, with the survivors passing on their successful traits.

The neural architecture deserves mention. Agents use NEAT brains that can grow new neurons, prune connections, and evolve among 7 different activation functions per node. The system includes an optional "cortex" - a fast associative reflex system that agents can evolve to rely on or ignore entirely. This cortex_reliance trait is itself evolvable, meaning some lineages might become pure NEAT-brain creatures while others maintain the cortex as backup.

Everything in Werld runs locally with Python 3.10+ and only the standard library - no pip installs required. The simulation auto-saves checkpoints and can be resumed or run for specific tick counts. For those who want to watch the evolution unfold, there's a separate dashboard built with Next.js and Recharts that polls a SQLite database every 4 seconds to show population dynamics, brain topology, species trees, and individual agent inspection.

The project's architecture is worth understanding for anyone interested in building similar systems. The simulation is split into two main parts: the pure Python simulation engine and the observatory dashboard. The engine handles the core loop - perceive (BFS neighborhood scan) -> decide (NEAT forward pass) -> act (continuous effectors) -> learn (cortex reinforcement + memory). Reproduction uses sexual crossover with NEAT gene alignment, allowing for genetic diversity.

For developers interested in contributing or experimenting, the project structure is well-organized with clear separation of concerns. The config.py file contains all tunable parameters, making it easy to tweak the simulation without diving into the core code. The CLAUDE.md file provides a comprehensive technical reference covering everything from sensory channels to genome traits.

Werld represents a fascinating approach to artificial life and AI development. By removing human-imposed goals and letting agents evolve their own intelligence, it creates a space for truly open-ended emergence. The fact that agents can discover new sensory channels, develop communication systems, and create complex behaviors without any programmer intervention is a powerful demonstration of what's possible when we let evolution do the work. Whether it leads to anything resembling human intelligence remains to be seen, but the journey of discovery is already proving fascinating.

The project is open source under the MIT license, and the creator is actively seeking contributors. If you're interested in artificial life, evolutionary computation, or just want to watch digital organisms evolve in real-time, Werld offers a unique and accessible platform for exploration. The combination of local-only operation, comprehensive observability, and open-ended design makes it an ideal playground for anyone curious about the fundamentals of intelligence and evolution.