The Four Generations of AI Coding

The first time I used AI for coding, it was with ChatGPT. Useful, but frustrating. Every session started from scratch. I had to re-explain the problem, repeat context, remind it what we were doing. One lapse and it would veer off completely.

It felt like pair programming with someone who had amnesia.

Then I tried Cursor. The AI could read the codebase, understand the structure, and help implement features without me narrating every step. What used to need paragraphs became a one-liner. A year in, I used it to migrate a Vue 2 project to Vue 3. Weeks of manual work, done in a few days.

I also built an AI-assisted workflow in VS Code GitHub Copilot and trained engineers on it. Teams saw around a 50% boost in development speed. It became something worth systematising.

I then explored OpenHands, where an agent writes code, executes it, runs tests, and iterates in its own environment. Then Claude Code, pursuing a similar direction through command-line workflows.

Looking Back: Three Generations

Looking back, I see three generations:

Gen 1 — Chat-based assistance. AI answers questions but the developer carries all context. (ChatGPT)

Gen 2 — Context-aware IDE copilots. AI reads the codebase and understands project structure. (Cursor, VS Code - GitHub Copilot)

Gen 3 — Autonomous coding agents. AI runs commands, tests code, iterates. The developer directs and reviews. (OpenHands, Claude Code, GitHub Copilot, Antigravity)

What's Gen 4?

Gen 4 — Goal-directed software systems (prediction)

The shift isn't about smarter agents. It's about who sets the agenda.

In Gen 3, a developer says: "Refactor this" or "Build this feature." The agent executes.

In Gen 4, a developer says: "Improve retention" and the system figures out what to build, builds it, measures results, and iterates. The developer reviews outcomes, not implementations.

These systems would maintain persistent context. A living memory of the product's history, architecture decisions, what was tried and why it failed. Like a senior engineer who never forgets anything.

Multiple agents would work in parallel: analysing behaviour, proposing solutions, implementing, reviewing, monitoring. The developer shifts from being inside that process to overseeing it.

At the far end, a product runs as a near-autonomous loop — identifying weak points, shipping fixes, adapting to users without waiting to be told. Not a coding tool. A product team that never sleeps.

The Pace of Change

What's strange is how fast this is moving. Six-month cycles now feel like weeks, partly because AI is accelerating the development of AI tools. The feedback loop is compounding.

The Shifting Role

The role is shifting. Less like craftspeople building line by line, more like technical directors shaping what gets built and why.

The skill that matters most isn't writing code. It's judgment — knowing which goals are worth pursuing, which tradeoffs are acceptable, and when to override the machine.