Architecture

Architecture articles

How the pieces fit once one prompt is not enough: splitting work across models, holding state between steps, orchestrating a run that can fail halfway, and resuming it without paying for the expensive half twice. Written around the tradeoffs, because most of these decisions are only reversible at cost.

All Fundamentals Building Architecture Operations Go-to-Market Monetization

Design the Human Handoff as a SystemEscalation gets referenced everywhere as a checkpoint and designed almost nowhere. The trigger, the payload, the queue, and the return path.4 min
Prompt Injection and Agent SecurityWhy an agent treats a malicious document as an instruction, and which defences actually hold when it does.6 min
Design Skills That Compose Without CollidingOnce you have several skills, the failure mode changes: the wrong one fires. How to draw boundaries and chain skills deliberately.3 min
RAG Deep Dive: Retrieval That Actually WorksChunking, embeddings, hybrid search, and reranking for domain-specific agent knowledge.3 min