AI amplifies what you already have — and other lessons from VP feeds this week

The most honest observation circulating in senior engineering circles this week came not from a keynote stage but from Gergely Orosz, who writes The Pragmatic Engineer and has spent years covering VP-level decisions at big tech. He quoted the founder of an AI coding harness admitting: "We're shipping way more hacks where we should have just rethought the whole system from the ground up. So I think our judgement is just off."1

That admission traveled fast because it put words to something many engineering leaders have been feeling. Orosz followed it with a sharper structural claim: AI amplifies good engineering cultures, and bad ones.2 Which made the obvious question louder — so why are so many companies gutting the middle management that maintains culture? He called it irrational given the moment.2

Those two posts, read together, describe the structural bet every FAANG VP is now making whether they say so explicitly or not.

Jaana Dogan, a principal engineer at Google focused on distributed systems and agentic workloads, announced the open-source release of Agent Executor (AX), described as a general-purpose runtime for agentic systems addressing dynamic scheduling, resumption, auto-recovery, auditing, and trajectory branching from kernel snapshots.3 The project is positioned as Google's answer to the infrastructure question that shipping more agents keeps surfacing: who coordinates the agent, handles failure recovery, and maintains an audit trail?

What makes AX notable as a signal is the language choice. Dogan had noted a few days earlier that Go has become the de facto language for agentic systems at Google, a development she called surprising given Go's historical position relative to C++. "I've never seen a path for Go to take over C++, but now I believe it is possible."4

Her critique of the current coding-agent moment runs parallel to the Orosz thread: "What's happening right now is that the coding agents are promoting feature maximalism and making it easier to introduce paper cuts. Lower layers of the stack need opinionated design, minimalism, and robustness."5 Two senior engineers, different companies, same week, same diagnosis — AI tooling is enabling more surface area, not necessarily better architecture.

On the model side, Jeff Dean, Chief Scientist at Google DeepMind and Gemini lead, promoted Gemma 4 12B — a multimodal open-weights model that runs directly on a laptop, released under Apache 2.0.6 The framing was deliberate: a "super capable" model for consumer hardware closes a gap that had kept agentic experimentation anchored to cloud APIs. For early-career engineers watching the VP feed, this is a signal that Google is competing in the on-device and open-weights layer, not just the API layer.

Dean had also participated in a recorded conversation with Oriol Vinyals and Noam Shazeer on Gemini's direction, amplifying the message that speed-capability tradeoff at the Flash tier is the current strategic focus.7

While Google was releasing open infrastructure, Meta was making its enterprise ambitions explicit. Naomi Gleit, Meta's Head of Product and the executive leading the company's new Enterprise Solutions team, announced the global launch of Meta Business Agent — an AI agent available on WhatsApp, Instagram, and Messenger that can book appointments, close sales, answer questions, process payments, and route complex queries to human staff.8

The quote that summarizes Gleit's strategy: "The number one thing I hear, especially from small businesses, is 'I just want to go to one place that can do all the things.' You want to make things modular, and you also want to be willing to evolve, because the technology is moving so quickly."8 The launch also included a Business Agent Platform connecting hundreds of third-party systems including Shopify and Zendesk — essentially a horizontal integration play across small business software.

This is a different attack surface than Google's developer-focused open-source runtime or Amazon's bedrock infrastructure bets. Meta is capturing enterprise AI by owning the communication channel where 100+ million businesses already operate.

Byron Cook, VP at Amazon, gave a talk at Columbia Engineering this week titled "The Fundamentals Behind the Future of AI."9 The core argument, per the Columbia summary: AI makes the fundamentals of computing more important than ever, not less.

This runs counter to a common framing in the industry — that AI-assisted coding reduces the need for deep CS foundations. Cook's position is the inverse: when systems are more complex and AI handles more implementation detail, your ability to reason about correctness, complexity, and system behavior at the foundation level becomes a sharper differentiator. It's a view that resonates with what both Dogan and Orosz were saying from different angles: AI amplifies, it doesn't substitute.

Three of these five voices are saying some version of the same thing: AI doesn't replace good judgment, it makes its absence more consequential. The two companies (Google and Meta) are executing that belief in opposite directions — Google building open developer infrastructure, Meta consolidating enterprise distribution.

For early-career engineers, the actionable read is this: the VPs who are winning right now are not the ones who treat AI tooling as a shortcut to better architecture. They're the ones who treat it as a multiplier on already-sound systems thinking — and they're staffing accordingly.