Experience evaluating and observing agent/LLM systems - building eval sets, defining success criteria, and instrumenting tracing, monitoring, and drift detection (using tooling such as Datadog, LangSmith, or Langfuse) so you can point to numbers that show whether a system is actually working. Design agentic systems deliberately: decompose the task, define the agent''s goals, tools, and action space, choose between single-agent and multi-agent orchestration, and specify how it plans, reasons, retains memory/state, escalates to a human, and stays within its guardrails.