Gary

Great post, Anshuman. Two decades in the industry here too, and I want to echo a few things and push back gently on one. On tests as spec — completely agree, and it's a powerful insight. But I'd add: the real leverage for AI agents isn't in rewriting existing logic. It's in integration and new features. And for that, you still need someone who understands the architecture well enough to make extension feel natural rather than bolted on. Tests tell the agent what, but architecture tells it where and how. Both matter. On "vibecoder" — I'd argue we've already moved past 1.0. Vibecoding was the prompt-and-hope era. What we're doing now is closer to agentic engineering: structured context, clear rules, human-in-the-loop review. Different discipline entirely. On AI slop — I love the evolution analogy, but I'd frame the slop problem slightly differently. LLMs generate from abstraction, not from a grounded world model. They're pattern-matching against our prompts without the real-world reference humans naturally carry. That's why the filtering mechanism — your test suite, your human review — is doing so much heavy lifting. As action agents mature (physical-world grounding, embodied feedback loops), I think this gets meaningfully better. For now, good prompts, explicit rules, and experienced humans in the loop are what separates signal from noise. On where this goes — fully agree that the architect role expands. I'd go further: until agents have a genuine world model, there will always be a seat at the table for experienced humans. Not just for catastrophic failures, but for the quiet judgment calls that happen a hundred times a day. And yes — AI agentic coding is real and here to stay. I've shipped several internal tools fully built by agents: a new language/framework scoping tool, a description review and improvement system, a next-gen SAST engine prototype. All in production, all generating real value. The secret sauce in every case was solid engineering principles underneath — SOLID, DDD, TDD — giving the agent a structure it could reason about and extend cleanly. The tools have changed. Good software engineering hasn't.