Darwin Agent: The What and When of Self-Evolving Agents
This article argues that self-evolving agents should be engineered as systems, not prompted as one-off models. It walks through the harness as a first-class object, memory as a lifecycle contract, world models as decision-support components, and multi-agent interaction as a topology problem — in a more formal, publication-style layout.