Writing from Luiz Ozorio on AI systems, agent architecture, LLM infrastructure, observability, evals, and supervised control systems for reliable intelligent software.
There's a quiet assumption underneath most AI discussions: if capability keeps improving, coherence will eventually follow. Gödel's incompleteness theorems suggest otherwise. The supervision layer isn't a temporary workaround — it's part of the design.
Once you can see what the system is doing, the next problem is interacting with it. Chat interfaces for AI agents are like flying a drone through a text terminal. The industry needs purpose-built control surfaces.
AI systems fail in ways that look like success. You can't find these failures in a chat log. You need traces. AI agent systems need the same observability infrastructure that distributed systems built over the past decade.
We're building AI systems like they're chatbots. They're not. They're control systems. The architecture that robotics solved decades ago — machine, telemetry, interface, human — is the same architecture AI agents need.
Most AI pipelines have generation and verification backwards. Smart models should generate. Cheap models should verify. The industry is putting its best capability in the wrong place — and it's creating a quality ceiling, not just a cost problem.
The industry phrase 'bash is all you need' is technically correct. But it hides the real problems: context management, execution reliability, alignment with outcomes, and human supervision.