Human Factors & the Agent-Era Career
The techniques above make you faster. This chapter is about what they can quietly cost you, and how not to pay it. The honest position, held across the sources and worth adopting, is that the productivity is real and the erosion is real, and the difference between the two outcomes is almost entirely about posture.
Three debts you can't see on a burndown chart
The agent era introduces failure modes that don't show up as bugs. It's useful to hold them as a trio, because they compound:
- Intent debt (Context, Skills & Memory): the why was never written down, and no agent can regenerate it.
- Comprehension debt (Osmani): "the growing gap between how much code exists in your system and how much of it any human being genuinely understands." It accrues invisibly (the code is clean, the tests pass) while understanding erodes, because AI generates faster than anyone can evaluate. Osmani cites research finding delegation-heavy developers scoring materially lower on comprehension of their own systems, and the memorable frame is "making code cheap to generate doesn't make understanding cheap to skip."
- Cognitive surrender (Osmani): the acute version, accepting AI output as your own without forming an independent judgment. It's distinct from healthy offloading (delegating the "how" while keeping the "what"). Surrender abandons the reasoning entirely. The cited Wharton finding is sobering: on the trials where the AI was wrong, people accepted the wrong answer 73% of the time, even as having AI available pushed their confidence up. You "borrow the model's confidence" without building a parallel understanding. Warning signs: approving PRs you didn't really read, accepting fixes without knowing the root cause, defending design decisions you can't actually defend, and lowered standards when tired. The line that locates the danger: "Surface correctness is not systemic correctness, and the gap between them is exactly where surrender hides."
What the research actually shows
The debts were named by practitioners. In 2025 researchers started measuring them, and the measurements land uncomfortably close to the warnings. MIT Media Lab's "Your Brain on ChatGPT" put essay writers under EEG across four months. The LLM-assisted group showed the weakest neural connectivity, the lowest self-reported ownership of their output, and frequently could not quote from essays they had "written" minutes earlier. When later forced to work unaided, their engagement stayed depressed, so the debt outlived the borrowing. (The caveats are a small preprint and essays rather than code, but it is the debt trio rendered in electrodes.) METR's RCT (Levels of AI Adoption) belongs on this list too. Developers 19% slower with AI while believing they were 20% faster is cognitive surrender's calibration failure measured in the field. And Microsoft Research and CMU's CHI 2025 study of 319 knowledge workers identifies the moderating variable this chapter turns on, that "higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking." The tool doesn't determine the outcome, which is Osmani's posture claim with a mechanism attached. The same study supplies the constructive reframe, that "GenAI shifts the nature of critical thinking toward information verification, response integration, and task stewardship." The thinking migrates rather than disappears, but only for the people who make the trip deliberately.
Don't outsource the learning
The most personal risk is to your own growth. Osmani's "Don't Outsource the Learning" makes the case that the default workflow optimizes for closing the issue, not building capability, and that the struggle you skip is the struggle that made you good. The striking datum he cites (an Anthropic 2026 study) is that AI-assisted engineers scored 50% on comprehension quizzes vs. 67% for those working manually, but, crucially, "the tool didn't determine the outcome. The posture did." Engineers who asked conceptual questions scored above 65%, while those who copy-pasted scored under 40%. The same tool produced opposite results.
The preserving practices are concrete and cheap. Form a hypothesis before prompting, ask for the explanation before the code, use a "learning mode" in unfamiliar territory, review AI output like a junior's PR, periodically recreate a solution by hand, and ask the model to teach its reasoning rather than just hand over the answer. His framing separates two metrics we tend to conflate. "Ship and learn are two separate metrics… I'd rather ship 80% of what I could have and learn 100% of what I needed to, than the reverse." That is a genuine trade-off, not a free lunch (sometimes shipping is the priority), but naming the two axes lets you choose deliberately instead of defaulting to ship-and-forget every time.
Charity Majors scales the same argument to the organization in "Generative AI is not going to build your engineering team for you". Engineering is an apprenticeship industry, "writing code is the easiest part of software engineering, and it's getting easier by the day," and what stays hard is owning systems in production over time, exactly the work agents don't do. Teams that stop hiring juniors because "AI does junior work now" are eating their seed corn. Comprehension debt has a generational form, where nobody is doing the reps that mint the next senior. Don't-outsource-the-learning is a hiring policy as much as a personal discipline.
Earning taste and judgment
The flip side of the debt story is where durable value now lives. Taste, the ability to look at a technically-correct solution and know it's the wrong one, doesn't come from reading agent output. It comes from the deep work agents are eager to do for you. This is the uncomfortable implication of cognitive surrender's cure, "mutual amplification, not delegation," with the goal of leaving a collaboration with sharper understanding, not fuzzier. Judgment is earned through the reps you're now tempted to skip, which is why the advice to "do the hard thing by hand sometimes" isn't nostalgia. It's how you keep the calibration that makes you worth having in the outer loop at all. The most credible optimists agree. Simon Willison (Here's how I use LLMs to help me write code) is blunt that the leverage is earned, warning that "if someone tells you that coding with LLMs is easy they are (probably unintentionally) misleading you." His own results lean "on 25+ years of professional coding experience," and his one non-negotiable is the outer loop itself, "the one thing you absolutely cannot outsource to the machine is testing that the code actually works." Expertise and AI leverage compound rather than compete. The taste you protect is what makes the same agent worth 10x in your hands and 1x in someone else's.
The new software lifecycle and the agent-era career
Stepping back, the shape of the work changes. The lifecycle reorders around specs, orchestration, and verification rather than authoring, so you spend less time in the editor and more time defining problems, dispatching agents, and judging results. Osmani's "The Agent-Era Career" states the thesis bluntly. "AI gets good at anything with an answer key. Your career is everything that doesn't have one." Problems with defined solutions get automated. Humans move to the ungradeable work.
Kent Beck, fifty years into his career, compressed the re-pricing into one line, that "the value of 90% of my skills just dropped to $0. The leverage for the remaining 10% went up 1000x" (90% of My Skills Are Now Worth $0). His reading of his own line is notably cheerful. Technological revolutions proceed by "radically reducing the cost of something that used to be expensive," then "discovering what is valuable about what has suddenly become cheap," and his prescription is aggressive experimentation rather than defensive abstention.
The value re-pricing is worth internalizing honestly. Falling in value are grinding boilerplate, raw coding speed, and solving well-defined technical puzzles. Rising are problem selection (deciding what to build and what not to), verification and quality judgment, deep cross-system understanding, spec-writing and clear thinking, and taste built through deliberate practice. The survival strategies follow. Do hard work on purpose (solve it without agents first to build the mental model), master verification ("there's nothing more demoralizing than delegation without verification at scale"), sprint the last mile (agents hit 70% fast, and the winning 30% is yours), own accountability (your name is on the change regardless of who typed it), and work in public (reputation is the scarce resource that compounds).
The honest trade-off underneath all of it is that efficiency gains genuinely risk atrophy of judgment, and output can look correct while your taste quietly degrades, which is why protection requires deliberate practice even when the agent could do it faster. None of this is an argument against agents. The productivity is real and the leverage is enormous. It is an argument for a specific stance toward them, the one Osmani keeps returning to and that this entire part endorses. Build the loop, run the factory, orchestrate the swarm, but do it as someone who intends to remain the engineer.