{"slug": "five-years-ago-i-made-a-prediction-about-code-heres-my-next-one", "title": "Five Years Ago, I Made a Prediction About code. Here’s My Next One", "summary": "Five years after predicting that developers would stop writing code by hand, a developer reports that over 95% of his code is now written by AI agents, with the shift moving from manual coding to 'harness engineering'—the system of constraints and verification around AI agents. OpenAI has described this shift, and Robert C. Martin, author of 'Clean Code,' advocates extreme constraints such as unit tests and mutation testing to validate agent-written code.", "body_md": "Back in 2021, I told a few colleagues that we would soon write almost no code by hand. GitHub Copilot had just arrived in VS Code and still felt mostly like improved autocomplete. It could suggest a few lines or an entire function, with results that were sometimes interesting and sometimes completely off the mark.\n\nMy colleagues had good reasons to be skeptical. Software development was too **complex**. Customers already struggled to explain what they needed. A machine would never be able to gather all the context required to build a real product.\n\nBut I was not predicting the end of code or the disappearance of developers. I only thought that **manually** translating an intention into thousands of lines typed into an editor would eventually be largely automated.\n\nFive years later, that has become the way I work. I am not telling this story to prove that I was right. I am telling it because I think a similar shift is beginning somewhere else.\n\nAfter code, software interfaces may be the next thing we no longer have to operate by hand.\n\nMost of the time, I begin by describing what I want to build out loud. I use **SuperWhisper** to turn my voice into text, then work with agents in **Codex**. I give them the requirement, the context, the constraints and, when necessary, precise technical guidance.\n\nThe **agents** then explore the project and modify the code directly. I can correct their approach, add a constraint or revisit a decision. I still occasionally open a file and make a change myself, but that is no longer at the center of my work as it was three years ago.\n\nThat figure of **more than 95%** is not based on a line-by-line audit, and my workflow does not represent the entire industry. It does show, however, that a different way of developing software is already possible.\n\n**The code still exists.** It still has to implement business rules, fit into an architecture and work in production. What has changed is the interface between my intention and that code.\n\nIn 2021, that interface was mainly my keyboard and editor. Today, I express an intention in natural language and agents work directly inside the project. I spend **less** time translating every decision into syntax and **more** time understanding the requirement, defining constraints and checking whether the result is actually the right one.\n\nCustomers have not suddenly learned to express their needs perfectly. If anything, that difficulty now matters even more. **Developers do not produce code for the sake of producing code.** **They use it to solve a product need.** When the mechanical part of the work can be delegated, the why and the what take up more space than the how.\n\nThis does not mean giving a model a vague prompt and trusting whatever comes back. The model is only one part of the system. Around it sits what is increasingly called a **harness**: the project instructions, tools, permissions, tests, linters and feedback loops that constrain the agent and make its work verifiable. OpenAI recently described this shift as harness engineering :\n\n[Harness engineering: leveraging Codex in an agent-first world](https://openai.com/index/harness-engineering/)\n\nRobert C. Martin (**Uncle Bob, a god in our industry**), the author of** Clean Code, **describes an even more radical approach. His current strategy is “not to read any of the code written by my agents.”\n\nInstead, he surrounds them with what he calls “extreme constraints”: unit tests, Gherkin tests, QA procedures, quality metrics, mutation testing and test coverage. His confidence does not come from blindly trusting the model, but from making its output run what he calls “the gauntlet” of those constraints.\n\nNot every team should necessarily go that far. But the principle matters: when writing code becomes cheap, engineering quality depends less on trusting the model and more on the system that constrains and verifies its output.\n\n**The work has not disappeared. It has moved.**\n\nThat shift made me look differently at every other piece of software I still had to operate through its interface.\n\nThe next generation of SaaS will be agent-first.\n\nFor the past five years, a large part of my work as a CTO has involved building SaaS products. I have watched their usage patterns and industry conventions evolve. I do not think these services will disappear. I do think their website or app will soon stop being the only way to access them.\n\nFive years obviously does not make me a software historian. But it is long enough to have spent a great deal of time designing interfaces, structuring business rules and observing how users actually access a service. I am beginning to see a separation between two things we have long treated as one: **the product itself and the interface we built to use it.**\n\nFor a long time, software had to be installed on the user’s computer. With the internet, browsers and sufficiently reliable connectivity, much of that software became remotely accessible services. Processing and data moved away from the user’s machine, but using the service still required opening an interface. First a website, and later often a mobile app.\n\n**The internet changed how software was distributed. Artificial intelligence could now change how we access and operate it.**\n\nSaaS products will keep their databases, authentication systems, permissions, business rules and ability to take action. But users will not necessarily need to open their interface for every operation.\n\nThat is what I mean by **agent-first **: a service designed to be used directly by an agent, without requiring its graphical interface **as the default point of entry**. This does not mean the interface disappears. It means the interface plays a different role.\n\nIt also requires us to design SaaS differently. Its capabilities must be understandable to a machine, its permissions explicit and its actions constrained enough to prevent an agent from confusing a read operation with an irreversible change. A good interface already protects users from certain mistakes. An agent-first service will need to provide comparable safeguards without relying solely on screens.\n\nThis is already how I use some tools. I can ask Codex to work with Buffer to organize and schedule my posts. My actual need is not to open Buffer, find the right screen and fill in several fields. My need is to publish a piece of content at a specific time, according to a set of rules.\n\nImagine I ask:\n\n> Schedule this article for next Wednesday at 9 a.m.\n\nThe agent might detect that another post is already scheduled for that time. To make a good decision, I would then need to see my other posts, their dates and perhaps their content. It could display a component containing my editorial calendar directly inside the conversation. I would look at it, then respond out loud:\n\n> Move this one to Thursday and keep the new one on Wednesday.\n\nI would never have opened Buffer’s website. Yet I would have used its service, its data and its business logic. A visual interface would still have appeared, but only at the moment it became useful.\n\nToday, interfaces are primarily used to execute actions: navigate, select, fill in, move and confirm. When agents handle that execution, the interface increasingly serves to show a state, explain a conflict, compare several possibilities and request a human decision.\n\n**Agents do not need an interface to act. We need one to understand and decide.**\n\nThe interface then becomes the place where humans review, adjust and approve what agents have done. It does not disappear from the software; it simply stops being the mandatory entrance to it.\n\nTo act without manipulating buttons or menus, an agent needs access to structured data and clearly defined actions. That is one of the purposes of the Model Context Protocol (MCP). Released by Anthropic in November 2024, it standardizes how AI applications connect to external data and tools.\n\n[Architecture overview - Model Context Protocol](https://modelcontextprotocol.io/docs/learn/architecture)\n\nMCP does not reason on the agent’s behalf. It allows the agent to discover a service’s capabilities and call them without going through its graphical interface.\n\nAnother protocol, Agent2Agent (A2A), addresses a complementary need. Created by Google and later donated to the Linux Foundation, it allows independent agents to discover one another’s capabilities and collaborate on a task.\n\nNow imagine a baker who works with several milk suppliers. The inventory system could detect that there is not enough milk left for the next day’s production. That event would *wake up* the baker’s agent.\n\nThe agent would use MCP to check the inventory and prepare an order. It could then communicate via A2A with the agents of several suppliers to retrieve their prices, available quantities and delivery times. The next morning, the baker might simply receive this message :\n\n> There is not enough milk in stock for tomorrow’s production. I have prepared an order with the least expensive supplier that can deliver before 11 a.m. Would you like me to place it ?\n\nAn interface could accompany the message to show quantities, prices and alternative options. The baker would only need to approve, modify or reject the proposal.\n\nThis example obviously assumes that inventory is tracked correctly, that suppliers expose accessible services and that the agent’s permissions are precisely defined. Artificial intelligence cannot magically guess what is inside a refrigerator. But the first building blocks already exist. Agents such as Hermes and OpenClaw can run scheduled tasks, react to external events and contact the user when something requires attention.\n\nThe appropriate degree of autonomy will naturally depend on the risk. Preparing an order, placing it and authorizing a payment are not the same action. The point is not to remove humans from the entire process, but to choose the exact moments where their attention and authorization **still add value**.\n\nIn this model, the user no longer initiates every interaction. A time, a change of state or an event received through a webhook can trigger the work. The agent proceeds on its own, then calls on the human when it reaches a limit or needs authorization.\n\nUntil now, we have opened software to ask it for something. Tomorrow, the software may come to us when a decision is still ours to make.\n\nI think we will increasingly use our voices to express intentions. Speaking is often more natural and faster than navigating through several menus or typing a detailed instruction. But there is no reason for the response to be audio-only.\n\nSome decisions require us to see a calendar, a table, a design, a chart or a code change. The agent can then generate or display the appropriate visual component. The conversation simply becomes the entry point from which several forms of interaction are possible.\n\nThat interface does not always need to appear on the same device, either. It could be displayed in a chat, on a phone, inside a work application or directly in the user’s field of view.\n\nThe Meta Ray-Ban Display glasses, released in September 2025, already offer an interesting signal. They include a color display and can be controlled with a wristband that reads muscle signals from subtle hand movements.\n\n[Meta Ray-Ban Display: AI Glasses With an EMG Wristband](https://about.fb.com/news/2025/09/meta-ray-ban-display-ai-glasses-emg-wristband/)\n\nDisplaying a Buffer calendar inside those glasses is not something you can do today. But it no longer feels speculative. The display technology already exists. What remains is largely a product integration problem: connecting agents, services and context in a way that is useful enough to become part of everyday life.\n\nInterface design will not disappear either. It will need to show what the agent understood, what it is about to do, the available alternatives and the consequences of approval. Designers may build more components that can appear in several contexts, across different devices and only when they are needed.\n\nWe will also continue to use full interfaces when we need to explore a large amount of information, manipulate a complex object directly or simply maintain an overview. Agent-first does not mean agent-only.\n\nThe more invisible execution becomes, the clearer those moments of review and approval need to be.\n\nOver a longer horizon, separating a reasonable prediction from science fiction becomes much harder. Interfaces could keep moving closer to us: glasses, wristbands, projections or devices we have not invented yet. Brain-computer interfaces represent another possibility, although a far more distant and sensitive one.\n\nNeuralink performed its first human implant in January 2024 as part of the PRIME Study. Its current purpose is medical: enabling people with paralysis to control external devices with their thoughts.\n\nThat does not mean we will be talking to artificial intelligence through our thoughts within a few years. The gap between an experimental medical device and a consumer interface is enormous. But the history of interfaces is unlikely to end with either the keyboard or the voice.\n\nThe specific device ultimately matters less than the overall direction: **we keep trying to reduce the distance between an intention and its execution.**\n\nMy prediction for the next five to ten years is therefore a measured one. SaaS products will not disappear, and we will not abandon all their websites and apps. But a significant share of our digital interactions could begin with an intention expressed to an agent rather than with opening a specific piece of software.\n\nThat agent will use several services in the background, sometimes in collaboration with other agents. Most operations will happen without a visible interface. When a comparison, an authorization or a human decision is required, an interface will appear in the context where we already are.\n\nIn 2021, I was not predicting the disappearance of code. I thought we would gradually stop typing it by hand.\n\nToday, I am not predicting the disappearance of software or its interfaces. I think SaaS products will become agent-first: their capabilities will remain available, but their app will no longer be the mandatory way to use them.\n\nThis shift has already begun in my own work. Its wider adoption is what I am now willing to bet on.\n\n[Five Years Ago, I Made a Prediction About code. Here’s My Next One](https://pub.towardsai.net/five-years-ago-i-made-a-prediction-about-code-heres-my-next-one-ee8fc43c56e0) was originally published in [Towards AI](https://pub.towardsai.net) on Medium, where people are continuing the conversation by highlighting and responding to this story.", "url": "https://wpnews.pro/news/five-years-ago-i-made-a-prediction-about-code-heres-my-next-one", "canonical_source": "https://pub.towardsai.net/five-years-ago-i-made-a-prediction-about-code-heres-my-next-one-ee8fc43c56e0?source=rss----98111c9905da---4", "published_at": "2026-08-13 12:31:01+00:00", "updated_at": "2026-08-13 12:50:04.284307+00:00", "lang": "en", "topics": ["artificial-intelligence", "generative-ai", "ai-agents", "developer-tools"], "entities": ["GitHub Copilot", "SuperWhisper", "Codex", "OpenAI", "Robert C. Martin"], "alternates": {"html": "https://wpnews.pro/news/five-years-ago-i-made-a-prediction-about-code-heres-my-next-one", "markdown": "https://wpnews.pro/news/five-years-ago-i-made-a-prediction-about-code-heres-my-next-one.md", "text": "https://wpnews.pro/news/five-years-ago-i-made-a-prediction-about-code-heres-my-next-one.txt", "jsonld": "https://wpnews.pro/news/five-years-ago-i-made-a-prediction-about-code-heres-my-next-one.jsonld"}}