Coding agent
Notes on my process
Building with AI
I learned to code without it, ignored it for years, then went all in. This is how I build now.
How I got here
- PDFs
Before AI
I learned to code before any of this existed. No assistants and not much material either, mostly PDFs I had to dig through on my own. Tough as a kid, but it taught me how to learn.
- GPT-3
First contact
GPT-3, while I was moving from React to Next.js. It invented imports and got the framework wrong. I decided AI could not code.
- Study
University
It earned a place as a study partner. Pass in a textbook, talk it through, go back and forth. Good for digesting material, nothing more.
Early 2025
Cursor started building whole features from plain English. Work that took me hours, done in minutes. I had ignored it too long, so I did what I do with any new skill: went all in.

Today
A local MCP server, a small software factory that automates most of my work, custom tools around my workflow, and voice agents.
Current stack
The subscriptions I'm running right now.
Coding agent
Codex
OpenCode with DeepSeek V4.1
Backup for token-heavy jobs.
Manuel MCP
A local MCP server holding the repeatable jobs I ask my agents to do. I built it because I needed it, not for anyone else.
Most of it exists as skills too. The MCP wins because any coding agent picks it up, whatever the model or the agent.
mcp
- Read my email
- Send email
- Automated emails
- Identity service lookups
- Speak out loud
- Snip and annotate
How it grows
- I ask an agent to do the job
- Slower, more tokens, but it gets done
- The job comes back again
- It becomes a tool in the MCP
Then the next job starts the loop again, and the server keeps getting bigger.
How my setup evolved
From a bare terminal to a bench of agents.
-
Then · The CLI
Claude Code
~ $ claude no CLAUDE.md no skills default everything > build the booking page Writing BookingPage.tsx > no, use the existing form Rewriting BookingPage.tsx > still wrong, start again ...
Claude Code in a bare terminal. Default settings, no skills, nothing configured. -
Next · Skills
- RPI Research, plan, implement
- TDD Tests first, then code
- Matt Pocock's skills
- My own skills Written as the work repeats
Then I learned about skills. My agents started following a process instead of improvising.
-
Now · T3 Code
T3 Code
Every agent, one window.
Claude Code Several subscriptions
Codex Several subscriptions
OpenCode Backup for token-heavy jobs
Orchestrator v2AutomationsManuel MCPVoice agentsSeveral Claude Code and Codex subscriptions run side by side, with OpenCode as the backup for token-heavy jobs. Orchestrator v2 coordinates them, and a lot of the routine runs as automations.