AI coding works in a sandbox because the environment is trivially narrow. Real systems have history, constraints, and blast radius. Coding agents make sound decisions only when the architecture is explicit and shared. Opinion isn't a constraint on agentic engineering, it's what makes it possible at scale.
A nature-restoration game with six biomes, 150 animals, and a real food web — built with a single Harper component as the entire backend. One YAML file wires the database, API, content seeder, and static host. The same binary ships offline on itch.io.
The web spent a decade optimizing for browsers. JavaScript-heavy rendering, dynamic CMS templates, and client-side hydration made pages beautiful and machines blind. AI answer engines retrieve, parse, and cite content directly. If your best content is trapped behind a render cycle, a cleaner source wins.
Harper 5.1 ships three layered AI capabilities: a provider-agnostic Models API supporting OpenAI, Anthropic, Bedrock, and Ollama; a built-in agentic loop via toolMode: 'auto' with hard budget controls; and an opt-in Harper Agent component. Switching providers is a config change, not a code change.
.jpg)
.png){kind=icon}
.png){kind=icon}
%20(1).png){kind=icon}