From Copilot to Conductor: My AI Development Journey

I didn't wake up one day and decide to become an "AI-first developer." It happened gradually, through a series of realisations that each changed how I worked. Looking back, there were four distinct phases - and understanding them might help you accelerate your own journey.

Phase 1: Autocomplete (2023)

Like most developers, my first taste of AI coding was GitHub Copilot. Tab-complete on steroids. It was impressive and useful, but it didn't change how I thought about development.

I was still:

• Opening files and editing them manually
• Running commands in the terminal myself
• Thinking in terms of "what do I type next"
• Treating AI as a faster keyboard

The mental model was "I write code, AI fills in the gaps." It made me maybe 20% faster, mostly on boilerplate.

Phase 2: Chat-Based Development (Early 2024)

The shift to chat-based tools - ChatGPT, Claude, and eventually Claude Code - changed the interaction model. Instead of AI completing my lines, I was describing what I wanted and getting back complete implementations.

This was transformative for certain tasks:

• Writing tests from a specification
• Converting designs into component code
• Debugging by pasting error messages and stack traces
• Learning new APIs and libraries

But I still had a one-at-a-time mindset. I'd ask for something, wait for the response, review it, apply it, then ask for the next thing. It was faster than writing everything myself, but the workflow was still sequential.

The mental model was "I describe what I want, AI writes it." Maybe 2x productivity.

Phase 3: Agent-Based Development (Late 2024)

When I started using Claude Code's ability to run background agents, everything shifted again. Instead of asking for one thing at a time, I could describe an entire feature and let multiple agents work on different parts simultaneously.

The key realisation was: I don't need to see the code being written in real-time. I can describe the outcome, set up verification, and let it happen in the background while I think about the next thing.

This phase also forced me to get much better at decomposition. You can't just say "build the user dashboard." You need to break it into independent, parallelisable pieces with clear interfaces between them. That's a skill, and it took practice.

The mental model was "I decompose work, agents execute it." Maybe 4-5x productivity, with some coordination overhead.

Phase 4: Team Lead Mode (2025 - Present)

The final leap was realising that even the coordination was something AI could do. I didn't need to manually create each agent and assign each task. I could have a "team lead" AI that handles all the coordination while I focus purely on direction and decisions.

This is where I am now. My workflow is:

1. I describe what I want at a high level
2. The team lead creates a plan, forms teams, assigns work
3. Agents execute in parallel, reporting back to the team lead
4. I monitor, redirect, and make judgement calls
5. Verification agents confirm everything works

The mental model is "I set direction, AI executes as a team." The productivity multiplier is hard to quantify because it's not just speed - it's a fundamentally different way of working.

What Changed at Each Phase

Each transition required letting go of something:

• Phase 1→2: Let go of writing every line yourself. Trust that describing the outcome is enough.
• Phase 2→3: Let go of watching code appear in real-time. Trust background execution.
• Phase 3→4: Let go of managing individual agents. Trust the coordination layer.

Each one felt uncomfortable at first. There's a very human desire to see the work happening, to feel the keystrokes, to maintain direct control. But each time I let go, I got back something more valuable: time to think.

Where I Think This Is Going

I don't think Phase 4 is the end. The next shift is probably autonomous goal pursuit - AI that not only executes but identifies what to work on next based on project priorities and user feedback. We're not quite there, but we're close.

What won't change is the need for someone with taste, judgement, and context to set direction. That's the one constant across every phase: the human decides what and why. The how keeps getting delegated further and further.

If you're in Phase 1 or 2 right now, I'd encourage you to experiment with Phase 3. Run one background agent on a small task. See how it feels. The transition isn't as scary as it seems - and you might be surprised how quickly your thinking adapts.