Built with AI under human supervision
Thinking log
Artificial intelligence has become an essential part of my daily workflow as a developer. Over the past few months I have used it to plan solutions, write code, solve problems, understand legacy systems, and automate repetitive tasks.
Concepts like SDD, Strict TDD and Agent Harnesses are changing how AI agents operate inside real systems. The focus starts moving away from isolated prompts and toward runtime discipline, persistent artifacts, validation layers and operational control.
AI coding tools can generate enormous amounts of code, but the deeper the workflow becomes, the more obvious it becomes that they still require constant supervision, reinforcement, and structured guidance to avoid making destructive or inconsistent decisions.
Accessibility tooling has improved significantly over the years, but workflows still require multiple tools, manual validation, and repetitive processes. a11y-engine started as an attempt to integrate AI directly into that workflow and make accessibility testing more useful inside modern frontend development.
Every time you use ChatGPT, search Google, or unlock your phone, a component is working silently behind the scenes: the microprocessor. This tiny chip is the engine that turns data into something useful — and it has become one of the most strategic assets on the planet.
One of the biggest problems with AI agents is not speed or code generation. It is memory. Most models constantly lose context, forget decisions, repeat mistakes, and require the same explanations over and over again. Memory layers are starting to change that.
AI can generate enormous amounts of code at incredible speed, but real development quickly exposes how unreliable it actually is. The deeper the project becomes, the more supervision, correction, and constant guidance the system starts to require.
Accessibility testing is part of modern frontend development. Tools like axe, Lighthouse, Playwright, and screen readers help validate how interfaces behave beyond visual implementation.
AI can generate styles, but it does not truly understand aesthetics, visual intention, or the design decisions behind an interface. In real projects, letting it modify styles often creates more inconsistency, noise, and cleanup work than doing the change manually.
Frontend development is no longer just about implementing layouts. In real projects, developers constantly face interfaces that ignore basic usability principles. Knowing how to adapt, question, and improve those experiences has become part of building modern products.
Working in technology makes it possible to observe how certain tools start integrating into real companies and eventually become infrastructure. Over the long term, many of the strongest companies are not the most popular ones of the moment, but the hardest ones to replace.
If AI already writes code, the question changes. The developer’s value stops being only about producing implementation and moves toward supervising, reviewing, validating, and being responsible for the real quality of the result.