Services
Most people in tech stop at ChatGPT.
That's not a criticism. The tools beyond it are genuinely harder to set up, harder to trust, and require a different mental model of what software can do. This is a map of that journey: where most people are, where they could be, and what it takes to cross each line.
Stage 1
The chat window
ChatGPT, Perplexity, Venice.ai, Claude.ai
Before
You open a browser tab, type a question, read the answer, and copy what's useful. It feels powerful at first. Then you notice you're doing the same thing fifty times a day, and the answers still need significant rework before they're useful to anyone.
After
You understand what the model is and isn't good for. You stop asking it to think for you and start using it to compress research time, draft first passes, and pressure-test your own reasoning. It becomes a fast, tireless junior: one you direct rather than defer to.
Most people stay here indefinitely. The jump to Stage 2 requires a change of environment, not just habit.
Stage 2
The AI-enabled workspace
Cursor, Kiro, AntiGravity, VS Code + Copilot, JetBrains AI
Before
You write code or documentation in one window and ask AI in another. Context switches constantly. You paste, copy, adjust, paste again. The AI helps but it doesn't know your codebase, your conventions, or what you were doing ten minutes ago.
After
The AI lives where the work lives. It knows your files, your patterns, your intent. Suggestions arrive before you finish the thought. Reviews happen inline. The gap between idea and implementation shrinks from hours to minutes.
This is where most engineers who consider themselves 'using AI' actually are. The next stage requires thinking differently about what a computer can be asked to do unsupervised.
Stage 3
Agentic workflows
Claude Code, OpenAI Codex, Google Gemini, Qwen Code
Before
You execute every step yourself. You read the error, form a theory, write the fix, run the test, repeat. You are the loop. Everything waits on you.
After
You describe the objective and the constraints. The agent reads the codebase, proposes a plan, executes it, hits errors, self-corrects, and surfaces a result for your review. You are the director. The loop runs without you in it.
Most people who try this stage abandon it quickly because their systems aren't built to be operated this way. Agentic workflows don't fail because of bad AI; they fail because of brittle architecture, missing observability, and insufficient trust boundaries. That's an engineering problem.
Stage 4
Production-grade agentic systems
MCP servers, tool-style APIs, hardened pipelines, custom agent frameworks
Before
You have something that works on your machine, in your repo, with your data. It is impressive in a demo. It would not survive contact with a real organisation: real users, real edge cases, real security requirements, real uptime expectations.
After
The system has defined tool boundaries. Agents can discover and interact with it safely. Failures are observable and recoverable. A senior engineer can read the code six months later and understand it. It ships. It stays shipped.
This is the last mile. It is where most AI projects die. It is also where two decades of systems engineering become relevant again.
If any of this feels like where you are right now, or where you want to be, I'm happy to talk through it.
No pitch. No methodology deck. Just a conversation.
Message on Signal