I sit on the bleeding edge of what AI can do every day. In my job at Google Cloud I work with dozens of customers every month. From fortune 50 to mid-sized to 10 person start-ups. If it’s interesting I’ll talk to them. Some are just starting their Agentic journey, trying to figure out how to translate their idea and their workflow into a collection of agents. Others are far more advanced, wiring up multi-agent orchestrations and reshaping entire delivery pipelines.
Most of them teach me something new along the way.
Here is what I see in the field: real teams figuring out how to integrate Agentic AI into practical workflows. Sometimes that is just a developer tweaking an inner loop to draft unit tests faster. Sometimes it is a complete reinvention of how a team approaches their job. Like an agent that independently drafts and verifies technical documentation. I see acceleration. I see setbacks. And I see genuine, steady growth.
Here is what I do not see: the breathless hype train that dominates my social feeds.
Throwing Out 20 Years of Instinct Is Not Progress
Multiple times a day, a post crosses my feed that makes me want to close my laptop and go build a birdhouse. Case in point:
A seasoned engineer claims that AI has forced him to abandon twenty years of software engineering instincts. He spent two decades removing duplicate code, building shared libraries, and preaching code reuse. Now? He tells his AI agents to do the exact opposite. It feels wrong, he admits, but he does it anyway because “that is the fun of engineering” and it lets him “become a beginner again.”
Let’s call this what it is: awful advice.
Twenty years of hard-won engineering discipline and architectural instinct are never invalidated by a tool. The Don’t Repeat Yourself (DRY) principle, modular design, and clean abstraction layers didn’t become standard practices because typing characters on a keyboard was hard. They became standard practices because maintaining, debugging, and reasoning about software over a five-year lifecycle is hard.
When you allow an AI agent to duplicate code across multiple files instead of creating a clean, shared library, you aren’t being forward-thinking. You are taking out a high-interest loan on comprehension debt. Sure, your agent generated the feature in forty seconds. But six months from now, when a critical edge case pops up or an API changes, you have sixteen slightly divergent copies of the same logic scattered throughout your codebase. Nothing fixes bad code.
The machine won’t suffer from that mess. Your users will. The craft of engineering is about managing complexity, not generating volume.
The “Software Factory” Is Not Here Yet
Another headline I keep seeing claims that the fully autonomous “software factory” has arrived. The narrative says you can fire your developers, hire three prompt writers, and let an army of autonomous agents build your core banking platform by Tuesday.
We’ll get closer over time, but right now that vision runs face-first into the cold laws of physics and economics:
- The physical supply of GPUs on planet Earth. Running complex, recursive agentic loops that read full codebases, reason across deep trees, and write diffs consumes massive compute. The data centers and silicon required to run entire software organizations purely on inference do not exist today.
- The current efficiency and cost of LLMs. Long-context inference remains computationally expensive and slow. Throwing hundreds of calls at every tiny bug is neither economically viable nor operationally practical for most organizations.
- The context window bottleneck. A context window is working memory, not an operating system. Until we establish reliable hybrid architectures where development state, workspace execution, and environment state live outside the model context, agentic development remains an augmentation, not a full reinvention.
There is fascinating research happening in this space. One project I watch closely is Scion, an open-source framework exploring agent harnesses and isolated execution environments. It tackles the hard plumbing: giving agents real workspaces and managing state where code actually runs. The architectural capabilities are beginning to emerge. The global capacity to run it everywhere at scale is simply not there yet.
Consciousness Does Not Live in Vector Math
Earlier this week, OpenAI launched GPT-6 Astra. Predictably, executives and tech influencers immediately declared that we have entered the “AGI era.”
I’ve written before about how the end of coding won’t be in 2026, and the same principle applies here. These models are amazing achievements in statistical inference and token prediction. Astra brings meaningful jumps in computer use and agentic workflows.
But it remains vector math.
Consciousness does not live in matrix multiplication. A model does not understand your customer’s unstated business requirements. It does not understand why your legacy billing system breaks when a timestamp lacks a timezone. It does not carry the burden of consequence. Treating a statistical model like an omniscient junior deity sets your team up for expensive failures.
The 2 AM Pager Test
I’ll be honest: sometimes the wave of hype gets so loud that I second-guess myself. I wonder if I’m just being stubborn. Has the world passed me by? Is it time for me to hang my sword over the mantle? Did I miss the revolution? The people building these models are exceptionally smart. They are doing brilliant, boundary-pushing research.
But interesting research does not always solve everyday business problems. And interesting things rarely make your pillow soft at 2:00 AM when you’re the engineer on-call holding the pager.
When production goes down in the dead of night, you don’t care how clever the prompt was that generated the broken microservice. You care about predictable behavior. You care about clear log traces, isolated failure domains, and well-understood dependencies. You want code that was designed with care, not code that was pasted together by an agent instructed to ignore twenty years of engineering wisdom. That is the world where intuitive leaps lead to innovation.
Agentic AI is a phenomenal power tool. I use it constantly, and I help customers harness it every single week. But it is an amplifier for your engineering judgment, not an excuse to abandon it.
Dodge the hype train. Keep your standards high. Focus on the solution.