AI in Coding, Part 5: How Far We've Come in Just 3 Years

Three years is not much in the history of software engineering. But in the world of AI, three years feels like an era. The difference between AI coding tools in 2022, 2024, and 2025 is like comparing a calculator to a smartphone.

Let’s look at how things have changed: and what that tells us about the pace of innovation.

2022: The Autocomplete Era

In 2022, most developers were just starting to experiment with AI in their workflow. GitHub Copilot had been launched a year earlier, and OpenAI’s Codex API was still in its early days.

Capabilities: Autocomplete for functions, comments-to-code, small code snippets.

Limitations: Weak context handling, frequent errors, no tests or security awareness.

Developer mindset: Curiosity. AI was a novelty, often used for side projects or experiments.

2024: Assistants Gain Traction

By 2024, AI assistants had moved from novelty to everyday tools for many engineers. Cursor IDE had launched, Replit Ghostwriter was maturing, and GPT-4 was setting new benchmarks.

Capabilities: Multi-file edits, better context awareness, basic refactoring, improved code explanations.

Limitations: Still weak on production readiness, limited support for CI/CD or testing.

Developer mindset: Adoption. AI became a regular part of workflows, especially for speeding up boilerplate.

2025: AI Agents Arrive

Now in 2025, we’re entering the era of agents. Instead of just suggesting lines of code, AI tools can propose pull requests, run tests in isolated sandboxes, and coordinate tasks across a repository.

Capabilities: Multi-agent collaboration, repo-wide edits, backlog execution, test writing, bug fixing.

Limitations: Still needs heavy human oversight, struggles with complex architecture, weak on deployment.

Developer mindset: Dependence. Many developers use AI daily and rely on it for productivity, but with caution.

Side-by-Side Comparison

Why This Evolution Matters

The last three years show us three things:

Pace of change is accelerating. What took a decade in earlier AI eras is happening in two or three years now.

AI is climbing the abstraction ladder. From autocomplete, to code assistants, to full repo-level agents.

Engineers need to adapt. The tools are getting more powerful, but also more complex. Knowing how to use them effectively is now part of being a good developer.

✅ Key Takeaway

AI in coding has gone from autocomplete curiosity to repo-wide agents in just three years. If this pace continues, the tools we’ll be using in 2027 may look as different from today’s as today’s look from 2022.

This article is part of my series “AI in Coding – From Past to Future.”

🎥 Watch the Full Lecture

For a deeper dive into this topic, watch my complete presentation on AI in software development:

Note: The introduction is in Bosnian, but the main lecture is in English and starts around the 3-minute mark.

This lecture covers all the topics in this series and provides additional insights and Q&A.

Full Series:

• Part 1: From Lisp to Copilot: The Evolution of AI in Software Development
• Part 2: AI Tools Every Engineer Should Know in 2025
• Part 3: Best Practices for Working with Your New Junior Developer
• Part 4: I Let AI Build BigToFit.com During My Lecture
• Part 5: How Far We’ve Come in Just 3 Years (this post)
• Part 6: The Next 10 Years of Coding (Coming October 16)
• Part 7: Will AI Replace Programmers? (Coming October 20)
• Part 8: Future-Proofing Your Career (Coming October 23)

What’s been your experience with AI coding tools over the past few years? Have you noticed this acceleration in capabilities in your own workflow? I’d love to hear about your AI coding journey and which tools have made the biggest impact for you.