Hi HN, I’m experimenting with a systems architecture rather than a finished product, and I’d really appreciate technical feedback from people working on ML systems, agents, or complex adaptive systems. The core idea is to treat strategies/agents as evolving “species” instead of static models. Each species is composed of modular “organs” (signal processors, risk controllers, execution logic, etc.), and mutations are constrained to compatible organ classes so components can recombine safely. Survival is determined by real-world performance under nonstationary environments rather than offline benchmarks. Instead of using LLMs to directly generate actions, I’m using them as an environment interpreter. A RAG layer ingests recent news headlines and other signals to classify the current “regime” or world-state, and species decide whether and how to act based on alignment with their intrinsic traits rather than explicit causal reasoning. This is very early-stage: – No production agents yet – Still wiring the architecture and data pipelines – Focused mainly on evolutionary dynamics and interface contracts I’m mainly looking for critique: similar systems you’ve seen, likely failure modes, scaling issues, or better abstractions. Happy to share diagrams or pseudocode if helpful. Thanks!