# AI Just Changed Who Gets to Be a Programmer > Source: > Published: 2026-08-11 04:11:34+00:00 For years, learning to code meant choosing a programming language. Python was often the first recommendation. It has readable syntax, a large collection of libraries, and strong support for automation, data science, and AI development. Python is still useful. Its popularity is not disappearing. But the most important programming skill is beginning to move one level higher. AI coding agents can now read a repository, edit several files, run tests, inspect errors, and revise their work. The human does not always write the code directly. The human explains what the software should do. That explanation is usually written in English. A vague instruction produces vague software. A missing requirement becomes a missing feature. An unclear definition of “done” gives the agent permission to stop at the wrong point. The person who can describe a system precisely may now have more control over the result than the person who simply remembers the most syntax. Python remains one of the most widely used programming languages. The Stack Overflow Developer Survey reported that Python adoption increased by seven percentage points between 2024 and 2025. It continues to be widely used for AI, data science, automation, and back-end development. GitHub’s data tells a slightly different story. According to the 2025 State of the Octoverse, TypeScript overtook Python and JavaScript to become the most-used language on GitHub by contributor count in August 2025. Python still added roughly 850,000 contributors during the measured period and remained the leading language for AI and data science projects. So Python is not dying. What has changed is the point where many people enter the programming process. Instead of opening a blank file and writing Python, they open an AI coding tool and describe the feature they want. The agent decides which files to inspect, what code to write, and which commands to run. The first language used is no longer Python. It is the language of the request. Traditional programming has a clear sequence: AI agents add another layer: Tools such as GitHub Copilot’s agent mode, Claude Code, and OpenAI Codex already work this way. GitHub describes Copilot agent mode as a system that receives natural-language instructions, examines a codebase, edits files, runs commands, checks errors, and revises its solution. OpenAI’s Codex can also read and edit repository files, run test suites, use linters, and submit changes for review. This is different from asking a chatbot to produce a small code snippet. A coding agent may work across an entire project. It can make dozens of decisions based on a few paragraphs of human instruction. That gives the original instruction much more influence. English is not executable in the same way as Python. A Python statement follows defined syntax. The interpreter does not guess what a variable “probably” means. It either processes the instruction or returns an error. Natural language is more flexible. If you tell a person to “make the page load faster,” they may ask what currently feels slow. They may check performance data, identify the main delay, and discuss what can safely be changed. An AI agent may choose one interpretation and start editing. The phrase “make it faster” could mean: Each interpretation can produce different code. When English controls an agent, ordinary words begin carrying technical consequences. Requirements, exceptions, permissions, and success conditions all affect what the agent builds. English does not replace code at the machine level. It becomes the instruction layer above it. A prompt asks for something. A specification defines what should happen. Consider this request: Add a login page to my app. The agent still needs to make many decisions. Should users log in with an email address or username? Does the app already have an authentication system? What happens after three failed attempts? Where should users be redirected? Does the page need password recovery? A more useful request would be: Add an email and password login page using the authentication service already configured in the repository. Place the page at /login. Validate the email format before submission and display the existing API error message when authentication fails. After a successful login, redirect the user to /dashboard. Do not add social login or change the database schema. Add tests for successful login, invalid email, and rejected credentials. This instruction gives the agent: The second request is longer, but length is not what makes it useful. It removes decisions the agent should not make alone. People often judge an AI coding tool by the code it produces. The real problem may have started before the code existed. If the instruction says, “Build a clean dashboard,” the agent has to guess what “clean” means. It may remove controls the user needs. It may hide data behind extra clicks. It may choose colors that do not meet accessibility requirements. Words such as “simple,” “modern,” “secure,” and “user-friendly” sound clear until someone has to turn them into behavior. Replace those words with conditions that can be checked. Instead of: Make the checkout more user-friendly. Try: Keep the checkout on one page. Preserve the customer’s entered information when payment fails. Show shipping costs before the payment step. Display validation errors beside the relevant field. Support keyboard navigation and keep existing mobile behavior unchanged. The improved version gives the agent less room to create an attractive solution that fails the actual task. Natural language has a serious weakness: it allows ambiguity. The same sentence can mean different things to different people. A requirement can sound complete while hiding several assumptions. Code is stricter. A function has defined inputs. A type system can reject the wrong value. A test can check whether an expected result occurred. A compiler can catch mistakes that a paragraph cannot. This is one reason typed languages may become more useful as AI writes more code. GitHub has argued that type systems provide a safety layer for code a developer did not write personally. Types can expose mismatched inputs, unexpected outputs, and incorrect assumptions before the application runs. AI agents may reduce the need to type every line manually. They increase the need for systems that can verify those lines. English can describe the goal. It cannot prove that the implementation is correct. AI coding agents make it easier to produce software without remembering every command. They do not remove the need to understand software. You still need enough knowledge to recognize: An agent can produce code that runs while still creating a security problem or a future maintenance issue. It may catch the visible error and miss the assumption behind it. This is why AI coding can feel easy at the beginning and difficult near the end. Producing the first version takes minutes. Deciding whether that version is safe to release requires judgment. The required skill is moving from code production toward code direction and review. The presence of an agent does not guarantee a productivity gain. METR conducted a randomized study involving 16 experienced developers completing 246 tasks in large open-source projects. In that setting, developers using AI tools took 19 percent longer to finish their assigned work. The developers expected AI to make them faster. Even after completing the study, they believed it had saved time. The researchers carefully limited their conclusion. The result does not show that AI slows every developer or every type of project. It applies to experienced contributors working on repositories they already understood well. You can read the full METR report here. The study highlights an important cost of agent-assisted coding. Generated code still needs to be read, tested, corrected, and understood. A weak instruction can create more review work than writing the change manually. English becomes valuable when it reduces uncertainty. More prompting alone does not improve the process. A coding agent needs a way to know whether it succeeded. “Make sure it works” is not a useful test. A clear acceptance criterion gives the agent something it can check: Tests turn part of the English request into an observable result. They also help when the agent misunderstands you. If the implementation violates a test, the failure creates immediate feedback. Without tests, a mistake may look acceptable until a user encounters it. OpenAI’s Codex documentation recommends configured development environments, clear project documentation, and reliable tests. Codex can run test suites and type checkers, but only if the repository provides them. The agent can use a safety net. Someone still has to build it. A strong agent request usually answers six questions. Describe the current behavior. The profile form loses all entered values when the server rejects the image upload. Define the desired behavior. Preserve the text fields and show the upload error above the image selector. Point it toward the relevant part of the project. The form is in src/components/ProfileForm.tsx, and the upload request is handled by src/api/profile.ts. Protect unrelated behavior. Do not change the API contract, route names, or existing validation messages. Give the agent acceptance criteria. Add a test that submits valid profile text with an oversized image. Confirm that the text remains in the form and the upload error appears. Request information needed for review. Summarize the files changed, explain the cause of the bug, and include the test command and result. These details make the work easier to review. They also reduce the chance that the agent fixes the visible symptom by changing something unrelated. Coding agents learn about a project by reading it. They inspect README files, tests, type definitions, configuration, comments, issue descriptions, and repository instructions. Some tools support dedicated instruction files. OpenAI Codex can use AGENTS.md files that explain the project structure, testing commands, and development rules. GitHub Copilot supports repository instructions for similar reasons. This changes the value of documentation. Documentation used to help the next developer understand the project. It now also affects how an agent edits that project. A poorly documented codebase forces the agent to infer more. Those guesses can appear in every change it makes. Clear documentation becomes operational. It shapes generated code. Beginners should still learn a programming language. Python remains a strong place to start because its syntax is readable and it allows new programmers to build useful projects quickly. But learning syntax alone is no longer enough. A beginner working with AI should also learn how to: Do not measure progress by how much code the agent produced. Measure it by how much of the result you can explain. If the agent creates 2,000 lines and you cannot describe the flow of the application, you do not control the project yet. Python communicates with an interpreter. English communicates intent to people and agents. Both can fail, but they fail in different ways. Python usually fails when its rules are broken. English can fail while sounding completely reasonable. That makes precise communication part of technical work. The strongest AI-assisted developers will not simply write the longest prompts. They will define problems clearly, place limits around the solution, provide relevant context, and create checks that expose wrong assumptions. They will know when to let an agent work and when to open the code themselves. Python is still important. But in a workflow where agents write more of the implementation, the first program may be the instruction you give them. And that program is written in English. [AI Just Changed Who Gets to Be a Programmer](https://blog.stackademic.com/ai-just-changed-who-gets-to-be-a-programmer-34f34d9a0440) was originally published in [Stackademic](https://blog.stackademic.com) on Medium, where people are continuing the conversation by highlighting and responding to this story.