After handing over 90% of the code to AI, what skills do humans still have?

How Cognition AI and its developed agent Devin are reshaping the future of software engineering. The author points out that AI has been able to take over 90% of the underlying execution work, including writing code and fixing bugs, liberating human engineers from trivial implementation details. Under this paradigm shift, the value of humans is redefined as core judgment abilities such as defining problems, designing architectures, and making decisions. By transforming the role of engineers into the "CTO of the agent team", AI greatly amplifies the output of talents with initiative, a global perspective, and tolerance for uncertainty. Ultimately, the moat of technological competition is no longer programming efficiency, but rather who can use AI tools to more precisely achieve business goals and innovation boundaries. This transformation not only significantly improves individual productivity but also foreshadows the arrival of the era of super individuals and micro - teams.

The engineers at Cognition AI rarely write code themselves anymore.

Because this work has been taken over by the AI they developed themselves. In March 2024, the company launched Devin, the world's first AI programming agent. The founder, Scott Wu, is a three - time gold medalist in the International Olympiad in Informatics (IOI) and won the world championship at the age of 17. From this top - level competition circle, a large number of talents who are reshaping the current AI industry have emerged: for example, Alexander Wang, the founder of Scale, Demi Guo, the founder of Pika, and Daniel Ziegler, the co - inventor of RLHF (Reinforcement Learning from Human Feedback).

In the past year, Devin has been successfully integrated into the real - world business environments of institutions such as Citibank, Santander Bank, and the US Treasury. In just the first two months of 2026, the amount of code delivered by Devin has exceeded the total for the entire year of 2025.

Now, engineers only need to spend 1 hour guiding Devin to produce the workload that used to take 6 to 12 hours.

When the core task of "writing code" is handed over to AI, where does the real value of humans lie?

Section 1 | Where Did the 90% Go?

Scott Wu said in an interview:

"In the past, when doing software development, about 10% of the time was spent thinking about what to do, and the remaining 90% of the time was spent dealing with implementation details."

This 10% includes understanding the problem, designing the solution, and deciding on the architecture; while the 90% is writing code, handling boundary cases, fixing bugs, making migrations, and implementing all the trivial execution steps.

Now, humans no longer need to do that 90% of the work.

Within Cognition, engineers no longer use writing code as their main way of working. They describe requirements in natural language, and the agent implements, modifies, and tests. One hour of work with Devin is equivalent to 6 to 12 hours of human work in the past.

Take product managers as an example. In the past, they needed to ask engineers: Why was this feature designed this way? What does that piece of code mean? Especially for new employees who were a bit nervous and didn't dare to ask these seemingly basic questions when they first joined the company.

Now they can directly ask Devin. Co - founder Russell Kaplan mentioned that Devin won't think your questions are stupid. It will answer any question you ask and even pull up the relevant code for you to see.

Moreover, the work that AI can do has far exceeded just answering questions.

There is an indicator for measuring AI programming ability called SWE - bench, which measures how long an AI can work on its own before human intervention is needed for correction.

Two or three years ago, this number was 10 seconds. After the AI wrote one line of code, the next line would be wrong.

Now, Claude Opus 4.6 can work continuously for 18 hours, equivalent to completing 18 hours of human work in the past, and no human intervention is required throughout the process. This ability increases by 4 to 5 times every year and doubles approximately every two or three months.

When humans no longer need to do that 90% of the work, the gap that used to be created by manual skills is being quickly narrowed.

Writing fast, writing accurately, and having a lot of experience, these past core competencies can all be achieved by AI. A person with little experience can also complete complex tasks with the help of AI as long as they can clearly describe the problem.

Russell also mentioned a detail: In terms of using AI tools, many people actually start at a similar level. The tools are updated every three months, and previous experience quickly becomes obsolete. In contrast, those who don't have a lot of old experience are more likely to adapt to this new way of working because they won't be stuck by old habits.

The most obvious gap between people in the past is disappearing.

Section 2 | What Is the Remaining 10%?

When humans no longer need to do that 90% of the work, what is the remaining 10%?

Scott Wu gave a very specific answer: understanding the problem, designing the solution, and finalizing the architecture.

In other words, before actually starting to type on the keyboard, engineers must first think through three things:

First, what is the optimal path to solve this problem?

Second, what is the most reasonable system architecture in the current scenario?

Third, what is the business goal I ultimately want to achieve?

In the past, clarifying these was just the first step, and there was a large amount of execution work to be done behind it. Now, the execution layer is completely handed over to AI. However, the new competition rule that follows is: when the tool threshold is leveled, whoever can more precisely define "what should be done" builds a real moat.

Russell Kaplan once worked on Tesla's autopilot team. A sentence from Elon Musk benefited him greatly: "Everyone is a Chief Engineer." It means that you can't just understand your own area of responsibility; you must understand how the entire system operates. In the autopilot team, there are multiple modules such as perception, planning, and control. To build a top - level system, you must have a precise understanding of each part and be well - versed in their collaborative mechanisms.

Russell believes that the arrival of the AI era has made this "global perspective" more important than ever. Because AI tends to connect the entire system for global optimization, the boundaries that used to separate different areas are now starting to blur. If you only understand one area, you can't judge whether the solution given by AI is reasonable.

Cognition's own recruitment logic is the best footnote to this trend.

Traditional technology companies often strictly prevent candidates from using AI during interviews, fearing it's "cheating." But Cognition does the opposite: The interview lasts for several hours, and candidates can use AI tools freely. The requirement is that candidates must build a complete product from scratch.

Within these few hours, it's impossible to finish the interview questions by writing code by hand alone. Being proficient in using AI has become a necessary condition. However, what the interviewers really examine is by no means "whether you can use the tool," but rather:

What kind of product do you think should be built?

How do you make product decisions during development?

How do you weigh and choose between different technical paths?

AI cannot answer these questions that require extremely high business and technical intuition.

The more accurately the problem is defined, the smoother the subsequent development and implementation will be. The more thoroughly the architecture is thought through, the better AI can execute in line with your intentions. However, it is still humans who judge "which solution is better" at countless crossroads.

Ultimately, AI has not made software engineering easier; it has just shifted the difficulty from "how to do it" to "what to do" and "how to choose."

Section 3 | Who Will Be Amplified?

This paradigm shift means not everyone can make a smooth transition.

After 90% of the basic execution work is stripped away, the ability gap between people not only doesn't narrow but is further widened. Here, AI plays the role of an "ability amplifier." It no longer amplifies traditional code execution ability but multiplies the following three traits:

First: Extreme initiative.

Scott Wu said that Cognition AI is essentially a "reasoning laboratory for building agents," and what they really value is employees' initiative and reasoning ability.

What is initiative? It means whether you can take the initiative to do things, not rely on a complete team configuration, and not wait for others to tell you what to do. Keep moving forward around the goal and use AI to complete the work that used to require a team.

In Russell's view, those with strong initiative have an advantage. They will think: "How much business impact can I make?" and drive the project forward on their own.

Second: Don't make choices.

A significant change is taking place among top - level engineers: they no longer struggle to choose between option A and option B.

In the past, limited by resources and energy, technical decisions were often one - way. But now, the best engineers will say: "Let's just run them all at the same time." They will break down the same difficult problem into multiple paths, assign them to multiple AI agents for parallel testing, and then choose the best result. In the same day, some people just use AI to complete one thing faster, while others explore five or six possibilities at the same time and constantly calibrate the direction.

The former improves single - line efficiency, while the latter broadens the innovation boundary.

Third: High tolerance for uncertainty.

Many engineering jobs used to be an extremely precise craft, and you had to control every detail. But now, you can't fully control what AI is doing, which makes people uncomfortable.

However, if you are willing to accept a little uncertainty, as long as you can ultimately verify whether the system meets the standards and accurately evaluate the output results, that's enough. The specific path to the result no longer needs to be closely monitored.

When Russell was developing autopilot at Tesla, there was a famous slogan: "Never go to sleep when the GPU is idle." The habit at that time was to start a large number of machine - learning experiments before going to bed and review the results when waking up. Now, he has translated this concept to all software development: "Why let Devin sit idle overnight? You can let it run batch tasks while you sleep."

When these three traits are combined, an individual's productivity will increase exponentially.

Take a large, strongly regulated enterprise as an example. They usually use tools such as SonarQube and Snyk for strict security vulnerability scanning. In the past, the massive amount of alarm logs made engineers exhausted. Now, they directly connect the alarm system to Devin for the first - round screening and automatic repair. The results show that Devin has successfully handled 70% of the vulnerability alerts in the production environment.

In this process, the positioning of human engineers has completely changed. Russell calls this new role the "CTO of the agent team." You only need to define the problem, then start your Devin army to work, and finally review the results.

According to Russell's judgment, super individuals and micro - teams are about to have a big explosion. Because when the execution cost is significantly reduced, what limits a person is no longer resources but whether they know what to do and can keep pushing things forward.

Facing this transformation, for those who already have a clear direction and the ability to act, AI will amplify these abilities several times. For those who rely on fixed processes and lack the ability to take the initiative, even with the same tools, it's difficult to have a qualitative change.

The ultimate gap no longer comes from the generation gap of tools but from the mindset of mastering the tools.

Conclusion | The Answer Is Simple

After handing over 90% of the work to AI, what skills do humans have left?

Three things: problem definition, architecture decision - making, and result selection.

These tasks that used to account for only 10% of the work are now everything. AI has not made people more equal. It amplifies the sense of direction, decision - making ability, and the ability to drive things forward.

Those who know what to do have an advantage.

Reference materials:

https://www.youtube.com/watch?v=-pZ3vD0r8a0

https://blog.joelonsdale.com/p/ep - 147 - scott - wu - and - russell - kaplan

Source: Official media/Online news

This article is from the WeChat official account "AI Deep Researcher", author: AI Deep Researcher, editor: Shen Si. Republished by 36Kr with permission.