AI: Back to Fundamentals (the ones you least expect) (TDC São Paulo)

TDC São Paulo 2026 (The Developer's Conference) ran September 23–25 at ProMagno, in person and online. My talk was part of the "AI & Emerging Tech" track—an English-language, international track—on September 25, alongside speakers like Edson Yanaga (Google), Csilla Szántó (MeteoSwiss), Nyah Macklin (Neo4j), Bruno Borges (Microsoft), Bharat Ahuja (Jeeves), and David Parry (Qodo). I also ran personal brand mentoring sessions at the conference's Mentoring Hub: Define your personal brand, or people will define it for you.

Summary of the talk AI: Back to Fundamentals (the ones you least expect)

When developers hear "fundamentals," they think of algorithms and data structures. This talk is about a different set: the software engineering fundamentals. The IEEE describes software engineering as a systematic, disciplined, quantifiable approach to building software, and the definition never mentions writing code.

As AI-generated output grows, our reviews get shallower. Uncle Bob says he no longer reads his agents' code. He trusts the constraints around them instead: an outer BDD loop (Spec → Gherkin) wrapping an inner TDD loop. That's the first fundamental to go back to. It's not new, just suddenly cheap.

The hole in the gauntlet

Every one of those checks asks whether the code does what the spec says. None of them asks whether the spec is right. When an agent writes the spec, the tests and the code, the judge is grading its own work, and you end up with a green suite and the wrong product. Reading didn't disappear; it moved up from the code to the spec. But we still read specs the old way: top to bottom, "looks fine to me," approve.

Perspective-Based Reading: the one you least expect

Perspective-Based Reading (Basili et al., NASA/GSFC, 1994–95) gives us an engineering way to read a spec. You don't just read it; you produce something from it: test cases as a tester, a design as a designer, documentation as a user. Reading lets you skip over a gap. Producing something makes the gap visible. In the demo, a three-sentence requirement that everyone agreed was fine raised 14 open questions, and an agent's Gherkin had silently "answered" every one of them.

What AI changes

PBR didn't fall out of use because it didn't work. It fell out of use because it was expensive: trained readers, hours each, and a meeting, for every requirement. AI makes it affordable, and it adds perspectives nobody could staff before, such as security, privacy or accessibility. Agents find ambiguity; humans resolve it. PBR becomes a gate at the spec level, before any code exists. The tooling is in pbr-for-specs and agentic-pbr.

AI didn't make the fundamentals obsolete. It made them affordable.