Agents write the code. I decide what ships.

AI agents write a lot of the code in my own projects now. Every change still goes through the same five steps, and the last one is always me.

1. An issue first. Anything a user could notice starts as a GitHub issue, so there's a record of why it exists.
2. An agent builds it on its own branch, often in its own worktree, so it can't trample work in progress.
3. It proves it works. Build and tests, and for anything visual, screenshots in a real browser at 320 px and on desktop.
4. A second opinion. In the pipeline that runs without me watching, a model from another vendor reviews the change before the pull request opens.
5. A pull request I read and merge. It closes the issue and carries its own changelog entry. Nothing lands on main any other way.

• A git CLI of my own that drafts branch names, commit messages and pull requests from the diff, with a local model on my Mac mini or a cloud model.
• MCP servers so agents work from real data instead of guesses: mulewatch reads MuleSoft logs, and my own apps expose theirs the same way.
• Meterlex tracks what my AI coding would cost at API prices. Building it is how I found that counting tokens from the logs overcounts by 2.5×.
• Local models on a Mac mini M4 Pro with 64 GB, for work that shouldn't leave the machine.

• Check what the model produced, not what it says it did. A text-to-speech model invented sentences in the audio of my notes, so every chunk is now transcribed with Whisper and re-recorded if it adds or drops a word.
• An empty review is not a pass. Reviewer models sometimes come back with nothing. The pipeline says so in the first line instead of calling the change clean.
• Agents don't invent facts. On this site, no figure, client or experience goes in unless I gave it, and client names never do.

This site is also an MCP server. Add https://api.heybrahim.com/mcp as a remote MCP server (Streamable HTTP, no key) in Claude, Cursor or any MCP client, and your agent can read my profile, availability, case studies, projects and notes, or search them.

It's read-only, with one exception: it can ask for my CV on your behalf, exactly like the form. I still read every request and send the CV myself.

GitHub Copilot every day (I'm GH-300 certified), plus Claude and Codex for review and generation. On my current MuleSoft mission I first built a log-reading MCP server for the client's estate, then released a client-free version as mulewatch.

What I'd bring to a team is the loop itself: agents doing the typing, proof before review, and a human who reads every merge. What I do now is on /now.