Will agentic engineering kill deep work? (part 2)
Summary
I continue to oppose the pernicious belief that agentic engineering diminishes the importance of deep work and that engineers and knowledge workers must now embrace a distracted, interrupted and context-switched workflow. Continuing where I left off in part one, I argue in part two that the above premise has more problems.
It ignores the sceptics
It ignores the AI backlash
It ignores the true costs of generative AI
In the previous article from this three-part series, I began fleshing out an argument against the following premise about agentic engineering.
“Future software builders and engineers must do less deep work and be ready for more interruptions as they orchestrate their agent army.”
The first three points in my argument were that the above premise:
doesn’t recognise the limitations of AI;
falls prey to marketing from the AI labs;
and oversimplifies the workflow from said marketing.
In this post, I’ll continue to explain why that assumption about agentic engineering is flawed.
4. It ignores the sceptics
Every story has two sides. Even if you love Boris and Andrej to bits, you must hear out the sceptics. Many senior developers have raised concerns about the agentic workflows that the AI labs want to sell us. Let me enumerate some of those concerns.
Using AI for any knowledge work, in this case, coding, takes judgment and taste. Taste doesn’t emerge from nowhere. Senior developers develop such taste through years of experience and hands-on coding. The less one codes, the more those skills atrophy, and the poorer one’s judgment and taste become. You needn’t believe me on this one. Even Anthropic acknowledges the phenomenon.
Junior engineers face this challenge even more. If you start your enterprise career pressing generate buttons, where does your judgment and taste come from? No wonder many junior engineers can ship fast but can’t debug issues in the hundreds of thousands of lines of code that an LLM spits out.
Margaret-Anne Storey has written about the idea of cognitive debt — where the codebase expands faster than the team can understand it. In a related post, James Shore writes about how coding agents can increase long-term maintenance costs. There is no world where humans won’t be responsible for software that goes into production, especially in industries with tight regulations. If the team can’t understand the code they ship, they risk introducing major vulnerabilities and creating a system they can’t maintain. After two major code leaks, perhaps even Anthropic will agree.
Of course, there could be ways to manage cognitive debt, but they need to slow down from the frenetic pace AI labs are marketing. Even Dax Raad, the creator of an open-source coding agent, advocates for slowing down.
“When working on something new or something challenging, me typing out code is the process by which I figure out what we should even be doing.”
Dax’s quote reminds me of the cognitive struggle that often precedes creative outputs. Typing on a blank page facilitates my thinking process. Thinking is the hard part, not typing speed. Similarly, many programmers write code as a precise way to think.
In an eloquent post, Lars Faye writes about all these perils of agentic coding. He thinks of LLMs as a powerhouse technological advancement, but suggests that we use them in secondary processes. In his words,
“I never ask an LLM or agent to implement something that I've never done before or couldn't do on my own, except perhaps purely for educational or tutorial purposes.”
Lars Faye is not an AI holdout. Nor is Dax Raad, or James Shore, or Margaret-Anne Storey. But they aren’t saying what executives want to hear.
5. It diminishes the AI backlash
It’s not that executives are unaware of these voices. They’ve heard them, read about them and dismissed them. Yet, companies and consulting firms are dismissing these arguments with several assertions that don’t hold up to scrutiny.
One such assertion is that senior engineers are Luddites experiencing a loss of identity, while junior engineers are AI-native and have embraced AI. To counter that assertion, look no further than the boos Eric Schmidt received during his tone-deaf prophecy about how AI will touch “every person and every relationship” people have. Who was booing him? Well, graduates from Arizona State University!
Futurism magazine summarises the state of AI backlash amongst younger people by quoting two important statistics:
Gen Z’s AI backlash relates to the broader anti-AI sentiment among people. You could argue that these sentiments exist only in the Global North. To some extent, countries like India indeed express less anti-AI sentiment, but that’s also because high-quality jobs have become scarce, even for the most educated graduates. Insecurity drives compliance.
A pointy-haired executive (not you, of course) could make the cynical choice to ignore advanced (more entitled?) economies and grow their agentic, engineering-compliant workforce only in the Global South. That could be a good thing, until you take the full costs into account.
6. It ignores the true costs of generative AI
Imagine a world where no one codes. That’s Boris’s world. Coding agents do all the work. They own the workflow. When the agent goes down, so does your workflow. Today, token costs are at the floor, even after Anthropic moved all enterprises to token-based pricing and GitHub Copilot switched its billing. Even if token costs go down, no one knows how to define a token. Claude Code decides how many tokens it’ll use to run your agentic workflow. If generative AI approximates a smart junior engineer (it doesn’t, btw), the labs decide how much that junior engineer will cost you.
In the last post, I already told you about Peter Steinberger’s 1.3 million dollar bill for three people. Take another example — Uber has burned all of its AI budget within four months. Individual engineers are spending anywhere between $500 to $2000 on tokens. In the world of tokenmaxxing, perhaps that’s OK for Uber, but is it OK for you?
“Maybe implicitly there’s more that is getting shipped, but it’s very hard to draw a line between one of those stats and ‘Okay now we’re actually producing like 25% more useful consumer features.’”
Suppose the labs are to be profitable (no one knows how), token prices must go up. Some companies will be able to afford the bait-and-switch. Others will need a hedge. Deep work and the continuing appetite for cognitive struggles could be part of that hedged strategy.
Next week, I’ll publish the final instalment in my thesis about agentic engineering and deep work. You’ll learn that all executives aren’t singing from the same song sheet and that speed isn’t always the bottleneck for software delivery. See you there.