IA and trades, episode 1.Computer developers now use assistants who can offer, modify and verify code. The global survey published on October 6, 2026 by Stack Overflow confirms this installation in daily work. But the field experiences and uses described by companies reveal a more precise transformation: delegating writing does not exempt from understanding the software or controlling its operation.
A standard language instruction can now initiate a correction, produce tests or prepare a program change. The developer is no longer content to write every instruction. It describes the expected outcome, provides context and examines a proposal. Agents can also chain several operations, instead of suggesting only a few lines in a publisher.
This is the first part of our weekly series on artificial intelligence and trades. It concerns a profession particularly exposed to these tools. However, it is neither a predicted disappearance nor an automatic increase in productivity.
Developers: broad adoption, conditional trust
In the 2026 survey, 65.9% of the 17,464 respondents to the usage question reported using programming assistants or agents. Generalist conversation tools collect 62.5%. Several responses were possible: these shares do not add up.
The size of the survey gives an indication of its size, not a guarantee of representativeness. It retained 30,903 valid responses from 169 countries, collected between 23 June and 5 August. Recruitment includes the site, blog and mailing list. These volunteers are therefore not a random sample of all developers.
Confidence remains primarily linked to the ability to control the response. Almost half of respondents say they trust AI when they can easily verify their work. Only 6.6% say they are willing to trust him for important decisions. The frequent use of the tool coexists with explicit limits to its autonomy.
This division of roles changes the nature of the work. Getting a function that seems correct becomes more accessible. Determining whether it respects product constraints, customer practices and project rules remains a separate transaction. A program that runs did not necessarily solve the right problem.
At Uber, agents enter the entire chain
The practices described by Uber show how far this delegation can go in a large undertaking. In a technical note dated August 27, 2026, the group assigns more than 70% of its requests for changes to local or cloud-hosted agents.
This figure does not mean that AI performs 70% of human work, nor does it write 70% of the code. A demand for integration may correspond to changes of very different size. The company mainly describes an extension of the operations entrusted to the agents: code review, automated test failure processing, alert analysis, bug correction and maintenance.
Human validations or referrals to engineers are still planned. The case thus illustrates an organization of work, with passages between automatic execution and human decision-making. However, these are results reported by Uber on its own systems, not an independent evaluation applicable to all enterprises.
Productivity depends on the work measured
The experimental studies give contrasting results. A paper published in February 2025 brings together three essays from Microsoft, Accenture and an unnamed large American company. Access to GitHub Copilot was randomly assigned to some of the 4,867 developers studied.
The combined estimate shows an increase of 26.08 % in weekly tasks completed among tool users. It takes into account that not all employees with access use it. Experiences last from two to eight months. Their separate results remain unclear, and activity indicators alone do not measure the quality of the software delivered.
Another experiment results in an inverse result in a narrower framework. Between February and June 2025, METR studied 16 experienced developers on 246 tasks. They work on large open source projects that they have known for five years on average. The authorisation to use AI increases their execution time by 19%.
The participants mainly employ Cursor Pro with Claude models of this period. This protocol therefore measures a demanding work done by specialists on their own environment. It does not describe all the developers, nor the performance of the agents available at the end of 2026.
The result also highlights the difficulty of judging its own time saving. After the experiment, participants still believe that the AI accelerated them, while the measures indicate a slowdown. The ease felt and the actual duration may differ.
These tests do not measure the same. One completed task account in several companies. The other chronometer of precise changes on familiar projects. Populations, tools and requirements differ. We can neither average their percentages nor choose a single number to describe the whole profession.
Tools that evolve faster than protocols
The further work of METR also calls for caution. On February 24, 2026, the organization explained that its new experience provides an unreliable signal. Some developers refuse to participate because they no longer want to work without AI. Others do not propose certain tasks to the protocol.
These choices are likely to rule out precisely where they expect the highest earnings. The simultaneous use of several agents also complicates the recording of time. New estimates suggest acceleration, but their uncertainty intervals include lack of effect.
The researchers consider it plausible to improve from the beginning of 2025, without being able to measure the extent of this with these data. Therefore, the initial slowdown cannot be transformed into a permanent verdict. Nor does it make it possible to announce that a universal gain would now be demonstrated.
Learning remains different from getting produced
Transformation also affects entry into the profession. An experiment published on January 29, 2026 recruited 52 engineers, mostly juniors. They know Python, but discover Trio, an asynchronous programming library. A party can use an assistant to perform two features, before an immediate questionnaire.
The average score in the assisted group was 50%, compared with 67% in the other group. The gap is 17 points. The observed slight gain in time is not statistically significant. This short experience does not demonstrate a lasting loss of skills, but distinguishes the production of a solution from its understanding.
Participants who ask for explanations or questions of substance understand better than those who delegate fully. However, this comparison between behaviours is not a separate randomized trial. It indicates a training path, without proving that a given method guarantees learning.
Employment is not derived from the number of lines generated
A research paper published on March 20, 2026 by two economists from the U.S. Federal Reserve examines occupations where programming plays an important role. He observed a much slower growth in employment since the arrival of ChatGPT at the end of 2022. However, the aggregate employment of these occupations continues to increase in the data studied.
The authors seek to distinguish this movement from the shocks of the employers sector. Their analysis remains observational. It does not transform every slowing down of recruitment into job deletion caused by AI. Nor do these American data measure the situation of developers in Lebanon.
The number of automated tasks is therefore not enough to count the jobs threatened or created. New projects, business needs and the distribution of work must also be observed. Available studies document some changes in tasks better than their sustainable translation into the workforce.
The correction no longer stops at the patch
At Shopify, a case published on September 2, 2026 illustrates this shift towards follow-up. Agent River checks a vulnerability in the current code, updates a correction and solicits engineers when a decision requires their context. After integration, it still controls the status of the repository and the dependencies of the software.
Shopify reports a decrease of about 70% in its stock of addiction-related incidents in eleven days. About two thirds of this decrease comes from integrated fixes. The remainder is for items that have become obsolete or already resolved. Therefore, it is not 70% of automatically corrected vulnerabilities.
This feedback has no control group. However, it describes a concrete task that goes beyond code generation: to follow a correction until the result is verified, then to involve an engineer when the risk or context imposes a decision.
Sources and references
- Stack Overflow, Developer Survey 2026: uses of AI (Stack Exchange Inc. data, ODBL 1.0 licence)
- Stack Overflow, survey methodology 2026
- Stack Overflow, results and confidence in AI, 6 October 2026
- Uber Engineering, agent-assisted development organization, August 27, 2026
- Cui and colleagues, three field experiments on developers, February 2025
- METR, Open Source Developer Experience, February–June 2025
- METR, limitations of the new experiment, 24 February 2026
- Anthropic, experience on programming learning, 29 January 2026
- Leland D. Crane and Paul E. Soto, AI and Coder Employment, 20 March 2026
- Shopify Engineering, River and Security Correction Tracking, September 2, 2026
- Illustration photograph: Kevin Ku / Pexels, graphic editing Libnanews
- Photo license Pexels
Comments