The Evolution of AI-Assisted Coding: A Personal Journey Through the Paradigm Shift
Last modified: April 14, 2025
AI-assisted software development landscape has changed on fast pace over the last two years (2023-2025) with most significant leaps over the last few months. The evolution is not just technological but can be described as paradigm shift how software engineering teams can work and what they can achieve.
I will try to illustrate this shift from personal angle, and provide a general timeline of these quite recent developments.
How I Met the Agents
In December 2024, while on vacation, I read a Twitter post
claiming that a new IDE called Cursor was pretty bad and
kept deleting users' files and code lines. Having done
plenty of coding with AI already, I wanted to see just how
bad it could be and installed it. Less than two weeks later, I had implemented a
feature-rich personal project with it, the term 'vibe-coding' had not
been invented yet, and my bones felt that the world had definitely changed.
Then January came, and I discovered Cursor's agent mode. We implemented the
game of Scrabble with the agent in just 40 minutes. Soon the agent took the lead
and starting to give instructions to me -- I had to create screenshots for
documentation it insisted we should write. Only a short while later I
realized that these "agents" were pretty good at using tools -- the command line
ones anyway. That happened as I was
updating a years-old experimental RAG/semantic search project (aiming it to use
local LLMs instead of OpenAPI's closed API). The agent created issues and pull
requests in GitHub and I had to admit that it was with greater clarity than
I would have managed myself.
(As a curious aside: the name "Cursor" seems to come from Latin,
with one meaning being "a slave who ran before the chariot of a grandee,
forerunner" – somewhat apt description for AI runnig ahead of us sometimes.)
Let's try to put some recent developments in a timeline. It is not
entirely accurate, but the idea is to illustrate the fast pace of change
and give some idea why I consider this a paradigm shift.
I'll make an effort to have some differentiation between personal
and general developments.
Lessons Learned and Drawbacks
Not everything has been smooth sailing of course. I've learned some painful lessons about not letting agents search for solutions from the wrong places, being more resolute about rolling back long chains of bad reasoning, and Claude Code taking all my money. When an agent starts down an unproductive path, it's often better to reset than to try to course-correct incrementally.
Challenges and Considerations
And there are some general challenges I want to mention:
- Addiction and learned helplessness: It's easy to become dependent on AI assistants and feel helpless when they turn their backs to you. When a model provider goes down during peak usage, it might become a blocker if you're in the middle of debugging a critical issue or implementing an urgent feature. One needs to remember where to fall back to.
- Tooling costs: The usage-based pricing of cutting-edge models adds up quickly when used intensively. You might need to keep a budget. The cost structure might create a barrier. (This is about Claude Code I guess.)
- Copyright uncertainty: I'm still not quite sure how (or if) the models steer clear of copyrighted material in the code and text they generate.
Anyways, I fully intend to blanket the internet with aggressively copyrighted Hello World programs.
- Geopolitical uncertainty/sovereignty concerns: We might have a good sense of where the tech is heading, but as for long-term access to our favorite tools — that may depend on which countries are still on speaking terms next quarter.
Where We're Heading Next
Looking ahead, I believe we should already be implementing agentic workflows for our clients' benefit.
We should have autonomous or scheduled agents doing busy work during nighttime hours,
agents that orchestrate and create other agents, and local LLMs taking on coordination of tools requiring higher standards for privacy.
Beyond the technology itself, the nature of organizations must start adapting to the arrival of a large number of artificial co-workers. I do not think it is just a change in tooling; it's rather a transformation in how teams are structured, responsibilities allocated, productivity achieved and communication arranged.
As we move forward, it seems natural that this transformation won't be limited to software development.
The patterns we're hoping to see in coding -- enhanced creativity, accelerated productivity, and human-AI
collaboration -- are likely to emerge in other creative and technical domains.
References