Prompt engineering is where most teams start and where a lot of them get stuck. It works until you have several domains, several agents, and a requirement that someone other than the author can understand what the system will do. This talk covers what replaced it in a system I built and now maintain in production. The core move is treating context as an engineering surface rather than a text problem: routing declared as configuration, memory and session scope defined explicitly, tool access permissioned so an agent's reach is a reviewable artifact rather than an emergent property of a long string. I'll cover the architecture concretely — hybrid rule-plus-LLM routing, DAG execution, session memory across multi-step jobs, retrieval scoped per tenant and per user, and page-level citation so a generated claim can be traced to its source. I'll also cover what this made possible operationally: quality gates, human-in-the-loop checkpoints, and per-call cost accounting that took worst-case job cost from over fifty cents to under five. I built this after working in LangChain and LlamaIndex and hitting their limits on real multi-domain workloads, so I'll be specific about where those limits are and when the added machinery is not worth it.

Target audience: AI/ML engineers, platform and infrastructure engineers, engineering leaders building agentic systems. Preferred duration: 30-45 minutes. Level: intermediate to advanced. Reference implementation is open source under MIT (GittieLabs AgentFlow, PyPI). No special technical requirements beyond a projector.