AI agents can generate and refine their own instructions using AGENTS.md, but effectiveness depends on clear human guidance, structured context, and iterative feedback rather than fully autonomous self-creation.
Harper v5.0 introduces a VM-based JavaScript environment that enhances application isolation, security, and developer experience. With application-specific context, module-level separation, and protections against prototype pollution, unauthorized access, and supply chain attacks, it delivers a more secure, scalable foundation for modern distributed applications.
AI agent costs are driven up by inefficient architecture. This guide breaks down five proven patterns, including deterministic workflows, parallel tool calls, and semantic caching, to reduce token usage, improve performance, and scale AI systems more efficiently.
An open-source real-time product recommendation engine built as a Harper component. Combines co-occurrence learning, HNSW vector search, UCB exploration, and category diversity re-ranking on Harper's replicated tables. No vector database, training pipeline, or external ML infrastructure required.
Build a real-time, geofenced promo engine on Harper's agentic runtime. The Nearstore Agent collapses geofence lookup, customer data, campaigns, and AI decisions into a single process. Clone the reference repo and deploy in minutes.
A Harper reference pattern: customer GPS ping → personalized promo decision via Claude Haiku, all in one Harper runtime. Geofenced with H3, data assembled in-process, pluggable decision layer.
.png){kind=icon}
.png){kind=icon}
%20(1).png){kind=icon}