Where Engineers Still Beat AI Agents, and How to Stay There

In 2026 it is normal for an AI agent to finish in an afternoon what used to take an engineer a month: a migration, a test suite, a working prototype. The trend is steep. METR, a research group that measures this, found that the length of software tasks frontier agents can complete on their own (measured by how long they take a human professional) doubled roughly every seven months from 2019 to 2025. Its January 2026 update puts the doubling time since 2023 closer to four months. So the useful question is not whether AI is fast. It is what is left for people, and how to stay good at it.

Start with what speed hides. In a 2025 randomised trial, METR gave 16 experienced open-source developers 246 real tasks in codebases they knew well. With AI tools they took 19% longer, yet afterwards they believed AI had made them 20% faster. The Stack Overflow 2025 developer survey shows the same tension: 84% of developers use or plan to use AI tools, but more of them distrust the output's accuracy (46%) than trust it (33%), and their top complaint is code that is 'almost right'. Agents are very good at producing something plausible. Someone still has to know whether it is correct.

Memory is the first place to compete, though not on raw recall. A model can hold a million tokens in context, more than most codebases, while human working memory holds about four chunks at a time. You will lose that contest. But an agent's memory is borrowed: it starts fresh each session unless someone writes things down, and it does not know why the payments service was split two years ago or which customer the last outage hurt. Human long-term memory turns experience into instinct: 'this looks like the race condition we had last spring.' To upskill, own the history. Keep short decision records, learn the business domain deeply, and explain systems to others, because explaining is how memory becomes judgement.

Reasoning is the second, and here it pays to be honest. ARC-AGI-2 launched in 2025 as a set of puzzles people solve easily and frontier models almost entirely failed; by mid-2026 top models were scoring at or above the average person. Do not build a career on puzzles AI cannot solve, because that list keeps shrinking. The lasting human edge is reasoning about the problem nobody handed you: which problem is worth solving, what the real constraints are, and what breaks when an assumption turns out to be wrong. An agent optimises the task you give it. Defining that task well is the hard part.

First principles are the third, and they are the cheapest way to catch a confident but wrong agent. Ask one for a fix to a slow service and it will happily add a cache or a retry loop. A minute of arithmetic tells you whether that can work. With Little's Law (in-flight work = arrival rate × time in system), 200 requests per second at 1 second each means 200 requests in flight; if your pool holds 100, no retry policy will save you, and retries add load. Maths, probability, complexity and queueing let you check an answer without having to trust it. As producing code gets cheaper, checking it becomes the scarce skill.

So is human intelligence degrading? The evidence points to skills fading from disuse, not to brains getting worse. In a 2025 MIT Media Lab study (a preprint, not yet peer-reviewed), 54 students wrote essays while wearing EEG caps. The group using ChatGPT showed the weakest brain connectivity, and in the first session 83% of them could not correctly quote a sentence from the essay they had just written, against about 11% in the other groups. A 2025 Microsoft and Carnegie Mellon survey of 319 knowledge workers found that the more people trusted AI on a task, the less critical thinking they reported putting into it. Over a longer span, Norwegian conscription data (Bratsberg and Rogeberg, 2018) shows IQ scores falling for people born after 1975, with the cause traced to environment rather than genes. None of this proves AI makes people less intelligent. It does show that what you stop practising, you lose, the same way GPS has made many of us worse at finding our way.

The fix is a set of habits. Attempt or at least sketch the solution before you prompt, so you have something to compare against. Read every diff the agent produces as if a junior engineer wrote it. Estimate before you measure: latency, cost, rows, users. Spend real time on fundamentals (maths, systems, and your business domain), because they are what make an answer checkable. Keep a decision log. And use AI as a sparring partner as well as a typist: ask it to attack your design, then decide yourself whether the attack holds.

AI is driving the cost of writing code towards zero. That moves the value to everything around the code: knowing what to build, remembering why things are the way they are, and being able to tell whether an answer is right. Don't try to beat agents on speed. Use them to go faster at things you understand, and never to skip the understanding.