From Copilot to AI agents: how my software development workflow has evolved
When I decided to become a developer, I thought I would spend my life writing code. Then AI agents came along.
To be fair, though, I don’t think they have taken my job, as I often hear people say. They have changed its shape, yes, but its essence has remained the same.
In this article, I share how my own evolution took place: from being “just” a developer to becoming something different, thanks to AI agents.
The evolution of my role as a developer
Today, in my daily workflow, I let agents handle a significant part of the code-writing process, but I don’t leave them to their own devices.
I step in to clarify the goal, come back often to adjust the direction, and check that the output meets expectations. It’s true: I write fewer lines of code. But knowing how to describe a problem, define a context, set constraints, and make sure the code is written as efficiently as possible is still programming work. In fact, it is a central part of the job: the part that separates those who truly program from those who simply write code.
From my perspective, using AI agents means changing how we look at the developer’s work. It means moving from being a code scribe to becoming an agentic software engineer: a developer who guides AI agents. By delegating part of the code-writing process, it becomes possible to free up time to explore larger codebases, reduce cognitive load, focus more clearly on solutions, and achieve results that would have been hard to imagine just a few years ago.
But this is only possible when agentic systems are truly integrated into your workflow, not just when you ask a GPT a question from time to time.
Working with agents, not just answers
When GPT models started becoming widely adopted three years ago, I too began asking GitHub Copilot for suggestions on how to work better. I would ask a question, the machine would suggest an answer, and I would apply it. It helped speed up the work, but that was not the real revolution AI was going to bring to software development.
Today, I no longer just ask it how to do something. I ask it to do it. I ask the machine to read a context, assess my goals, propose a plan, implement a solution, run checks and, when needed, explain some of its choices and correct itself.
How did i get here? The role of the community
At the 2025 Italian Agile Days, I attended a talk by Stefano Leli that presented a development experiment based on specialized agents. There were agents dedicated to product management, design, coding and testing, with real people responsible for reviewing each step.
That presentation opened my eyes and helped me see AI as an active element within the process, not just a chatbot to ask for help when needed.
A few months later, at XMAS Dev in Rome in December 2025, a presentation by Alberto Acerbis helped me bring other pieces into focus: Spec-Driven Development, context management and the use of markdown files as operational memory. The dots were starting to connect, and the wheels in my head were starting to turn.
The turning point: from software developer to “tech lead” for AI agents
At the beginning of 2026, the company I was working with decided to push strongly for AI adoption. I was given a Claude Code account and a very clear request: find a new way of working, one where GPT and people could reach goals together, faster and more efficiently.
Easy to say. A little less easy to do.
Bringing agents into your workflows first requires a shift in mindset. A developer’s instinct is to step in immediately, take control and fix what does not work. With agents, instead, you have to learn to resist that impulse and, when something does not work, stop and explain your request more clearly. This is exactly where my role started to change shape: from developer, I became the tech lead of a group of algorithms.
The agentic software engineer
Agents are tireless and fast collaborators, but they are not always reliable. To work well, they need the technical expertise of an experienced person who knows how to guide them in the right direction.
An agent can produce plausible code even when it has not truly understood the reason behind it. It can solve a narrow problem while ignoring a broader architectural constraint. That is why human oversight does not disappear. Using AI in an agentic rather than conversational way leads us, in a sense, to all become tech leads capable of guiding a team of machines toward a goal.
Agentic AI development and Agility
This is where agility, in the most concrete sense of iterative and incremental development, becomes essential again. With AI agents, I cannot hand a huge requirement over to the machine and expect the final result to be correct. I have to move in small steps, quickly surface a verifiable output, give feedback, adjust the direction and consolidate only what works. The faster AI becomes at producing output, the more important it becomes to shorten the cycle between request, result and validation.
The benefits of an agentic development mindset
By adopting an agentic mindset, you can also have multiple GPTs work on the same problem, activating different perspectives such as security, infrastructure impact and testability. In a short time, you can get a very accurate overview of the code and identify risks that a team of people would have needed full days of work to detect.
In real projects, software development is not just about writing code. You need to understand the domain, respect existing constraints, coordinate with other parts of the system, avoid regressions and maintain consistency with previous decisions. AI does not remove this complexity, but it helps make it more manageable.
In short, technical judgment remains in the programmer’s hands. Today, I am still the one who evaluates the output and decides which path to take, only faster and more safely.
Wiki-proof memory
Another thing I find revolutionary about using agents is their operational memory.
Despite some limitations, agents are able to build know-how that is recalled based on the tasks they need to perform, and that the agent updates automatically as it acquires new information.
This is a huge leap forward compared to the wikis we used to create with colleagues.
With AI agent skills, knowledge becomes more transparent and easier to replicate in the future. Inside the GPT, a reusable context is built: knowledge becomes explicit, and with a clearer workflow, the entire project is better organized and carries fewer risks.
The other side of the coin
Adopting an agentic workflow speeds up the work, reduces risk and makes outputs more accurate, but it also requires a good deal of awareness.
AI, Yes — But at what cost?
First of all, AI comes with costs that companies cannot underestimate. When agents are introduced into projects at scale, they also require a significant investment. It will be necessary to understand which activities should be assigned to more powerful models and which ones can instead be handled by local LLMs to keep costs under control.
Even with AI, expertise is essential
Then there is the expertise of the individual programmer to consider. Personally, I believe I have learned to use AI systematically, but other programmers are running more advanced experiments, such as agentic dashboards, personal assistants, and systems that autonomously generate tools, workflows and new skills.
Research is clearly moving in an agentic direction, but the future is still not fully defined, especially when it comes to concrete use cases.
My feeling is that we are still at the beginning: we have only started to understand where the agentic world can really take us and, as the curious person I am, I can only watch how my role will evolve over the next five years. One thing seems certain: as a programmer, I will not spend my whole life writing code.
Training the AI muscle
Today, when I face a problem, the first question I ask myself is: “How can AI help me?”. This is not laziness. AI is a tool, but it is also a skill that needs to be trained. You need to know the tools, from MCPs to skills and CLIs, and above all develop the habit of bringing them into your workflow, until it becomes a kind of muscle memory.
Using Claude Code, Codex, OpenCode, Pi, Cursor, Antigravity and the other agentic development tools has made me realize that there is still no clear winner in the race to become the best tool.
We are in the middle of the AI race and, as a developer, I find this one of the most interesting phases to go through. Seeing that it is already so effective while still having huge room for improvement, and seeing a community buzzing with enthusiasm, ideas and also a bit of healthy skepticism, makes me think that the most interesting part is still ahead of us.