Plan → act → observe loops with bounded autonomy: budgets, stop conditions, checkpoint gates, and human-in-the-loop escalation. Adversarially reviewed before it's trusted.

Deciding what the model sees on every call — system state, retrieved knowledge, tool definitions, memory. Most agent failures are context failures, and that's where I spend my token budget.

Multi-provider, tiered / cascade routing optimized for cost-per-successful-task, not cost-per-token.

Building Model Context Protocol servers and composable skills that give models governed access to real internal systems. Plumbing, not chatbots.

Hybrid RAG over messy real-world documents; vector search (Qdrant, pgvector) alongside open, markdown-based knowledge graphs the agents can read and maintain.

Golden sets from real inputs, regression suites, LLM-as-judge where it earns its keep. Observability without evals is just watching things break in higher resolution.