A conference talk covering engineering practices for building reliable, production-grade agentic AI systems. The core problem is the 'chaos marriage' of LLM stochasticity and distributed systems non-determinism. Three pillars are proposed: durable execution (event-sourced workflows that survive node failures and resume mid-flow), data intelligence (mining real-time signals like confidence scores and hallucination rates from event streams), and adaptive policies (deterministic code-level guardrails that constrain LLM outputs and can auto-route or self-heal). Additional patterns covered include domain engineering with bounded vocabularies, schema-enforced inputs/outputs, circuit breakers, human-in-the-loop escalation based on confidence thresholds, multi-agent adversarial debate, and compositional reasoning for complex tasks. The talk uses the Akka platform for demos.
Sort: