Control Systems for Intelligent Software

AI agents aren't chatbots. They're control systems — with the same architectural requirements as robotics and industrial automation: execution, verification, observability, control surfaces, and human supervision.

This is a system, not a collection of posts. Each layer builds on the previous.

strong model generates→constrained verification checks→observability instruments→control surface presents→human supervises

The Stack

  1. 1
    Execution

    Bash Is All You Need — Until It Isn't

    Bash is the perfect execution layer. But execution is one box in a much larger pipeline.

  2. 2
    Verification

    Why the Smart Model Reviewer Pattern Is Backwards

    Strong models should generate. Verification should be pushed toward constrained, cheaper, more deterministic mechanisms.

  3. 3
    Architecture

    AI Agents Are Control Systems

    The architecture that robotics solved decades ago is the same architecture AI agents need.

  4. 4
    Observability

    Observability for AI Agents

    AI systems fail in ways that look like success. You need traces, not chat logs.

  5. 5
    Interface

    Why AI Needs Control Surfaces, Not Just Chat

    Chat interfaces are a bottleneck. Purpose-built control surfaces turn telemetry into human decisions.

Boundary Condition

Why Coherence Doesn't Scale with Capability

Capability is scaling faster than coherence. The supervision layer isn't a workaround — it's part of the design.

Luiz Ozorio is a Staff AI Systems Engineer building production AI platforms — orchestration, evals, observability, and supervision for reliable LLM systems.

All writingAbout