Architecture 3.0: From 90% to 99.999% Reliability in Building AI Systems

We’re entering Software 3.0: not just AI inside products, but AI agents that design, implement, and evolve large parts of the system and increasingly build other agents. The hard question isn’t whether this works at ~90% “looks correct,” but how we engineer a bridge to 99.999% reliability when the workforce is probabilistic.

This talk presents an Architecture 3.0 blueprint for AI-native systems and agentic engineering pipelines: how to structure systems where agents generate code, configs, tests, docs, and even agent behaviors, while keeping control. You’ll see why the architect role shifts from diagram drawing to designing the control surface: specs that are testable, benchmarkable, and “reinforcement-compatible,” plus the thin layer of humanity focused on what creates long-term value.

DEMO (live): we let agents build a small agent-powered service and an internal “helper agent” (tooling/prompt + eval suite). CI gates intentionally fail on safety/reliability; we iterate on the spec + evals until the system passes with predictable behavior.