In the 50 years that I’ve been paid to write computer software, I’ve seen a number of technological revolutions. In 1981, I worked on developing software for the first IBM Personal Computer, even before it was announced to the public. In the 1990s, I watched as the graphical user interfaces of Microsoft Windows and the Apple Macintosh came to be the standard method of interacting with a computer. The Internet came along in the mid-90s and gradually connected things together. In the early 2000s, the web browser and web-based applications came to dominate software. Finally, smartphones and mobile applications ate the world, beginning with the iPhone in 2007.
Now we’re going through what appears to be the artificial intelligence (AI) revolution.
The world of software development is obsessed with the potential of AI to eliminate the need to write code. Hardly a day goes by when I don’t read a story about some software engineer who spends his time specifying what a program should do, rather than writing the code that does it. At the same time, I look at the work I do each day programming a computer game and wonder how AI would handle that. Specification can be more demanding than writing code, since English is an imperfect tool.
In fact, the company I work for recently held a session (via Zoom) entitled “GitHub Copilot Foundational Training,” intended to show software engineers how to write code using Copilot, Microsoft’s AI assistant (Microsoft also owns GitHub, which is where many companies store their software code). The final part of the session was intended to show Copilot extending an existing website that displayed product summaries.
Copilot, which had access to all the code for the website, was told simply to place a functioning “Add to Cart” button on each product summary. It added the button, but it was nonfunctional and there was no cart. When told this, it made another attempt, with much the same result. It was not an encouraging demonstration of the power of AI-based coding. Of course, it might have achieved better results if the specification had been more complete.
On the other hand, the general consensus is that the large language models (LLMs), which power assistants like Copilot, have undergone a step-change in capability over the past few months, particularly OpenAI’s GPT-5.2, Anthropic’s Claude Opus 4.5, and Google’s Gemini 3.
Unlike earlier versions, these models attempt to use reasoning in formulating their responses and can operate independently for long periods of time. This ability to operate independently is important, as it forms the basis for “agentic AI,” which IBM describes as “autonomous, goal-oriented systems that can independently plan, make decisions, and execute complex, multi-step tasks with minimal human oversight.” Basically, they’re replacements for human workers.
One AI tool company, Cursor, recently demonstrated how a swarm of AI agents (using GPT-5.2) worked autonomously for a week to produce a functional, albeit buggy, web browser at a cost of about $85,000 in computing resources. My favorite article title regarding the effort? “Cursor used agents to write a browser, proving AI can write shoddy code at scale,” from The Register (full article at tinyurl.com/5t5du4zy).
So, will AI agents replace software engineers?
If you believe the hype, in the very near future, product managers will specify what a piece of software should do and AI agents will go off and create it — at a fraction of today’s time and cost. It’s seductive to think that you could have a team of AI agents working for you, each as capable as an experienced human but costing less and working faster than the humans you presently employ. Bringing products to market faster and higher profits are music to the ears of any CEO or CFO. And yet the demo I watched this week indicates we are not yet there (and may never be).
Another problem: your competitors will have access to these same agents. If everyone can develop faster and cheaper, where does your competitive advantage lie? It’s still in the ability to come up with new and better ways to serve your customer, something that AI has yet to demonstrate. Even in the AI future, some humans will still be necessary, and their value to a company may greatly increase. Of course, there may also be a lot of people who are out of jobs. The AI future may not be a happy one.
A lot of this comes down to trust. Do you trust an AI agent to write bug-free code? Do you trust an AI agent to reorganize your files? Do you trust your AI agent to manage your calendar, respond to emails and make purchases? At the moment, it takes a certain amount of courage.
But, as someone has pointed out, this is the worst these tools will ever be, and they certainly already have the ability to assist humans in useful ways. Whether AI agents can operate independently, with human direction, at a high level remains to be seen.
