Marvin
Aziz
Marketer turned agentic builder, shipping software at an unreasonable pace.
Growth Engineer at Lindy in San Francisco, running product-led growth experiments. Built agents with thousands of professionals through a community we grew, before becoming what I like to call an "agentic builder" — a fancy way of saying I built myself a system with agents to ship software at an unreasonable pace. This site is where I write about it as I go.
ls -lt ./blog | head -5
- How I make videos with Claude Code~2308 tok · 9 min
- How to trust your agent~2447 tok · 10 min
- How I ship code I can't read~3271 tok · 13 min
- Why My Coding Agents Live on a VPS Now~1619 tok · 7 min
- The Four Layers of Agentic Engineering~3122 tok · 13 min
tail -n 4 ~/into.log
- [note]
My AI coding agent spend is getting ridiculous: 2x Codex Max, Claude Code Max, Cursor Max, OpenRouter, DeepSeek API, Baseten, and Cerebras.
- [gh]
Discovered Rustrak and now use it on my VPS. Already connected a bunch of my pages.
- [gh]
Installed bottom on the VPS. Rust-written htop replacement, runs right in the terminal. CPU, RAM, network, per-process, all on one screen. Finally know what is eating memory when an agent goes feral at 2am.
- [link]
Moved my coding agents off my laptop. They live on a Hetzner box now. Claude Code and Codex run inside tmux, reachable via mosh. I close the lid, walk to the café, reattach from my phone, they are still going. The work stopped being tied to where I am.
ls ~/projects
total 5- gettakes.app
takes2026The Mac app I built for everything after the recording: the cut, the captions, the posts.
- wpm.marvinaziz.de
wpm-screener2026Typing speed and accuracy gauntlet. Built as a hiring screen, then I got hooked.
- heynimo.com
nimo2026Website performance as a service. Subscribe to speed, not to charts.
- tennis.marvinaziz.de
sf-tennis2026Real-time availability map for public tennis and pickleball courts in SF.
- videobrainstorm.com
videobrainstorm2026One rough prompt becomes several image directions, then videos, finals and polls.
My coding-agent token usage
59.9B
tokens went through my coding agents in the last 30 days, across 7,096 sessions. This is who did the work.
58.2B of that is cache reads: context the agents re-read on every turn. 1.74B is new input, output and cache writes.
- gpt-6.1-sol58.0%
- claude-opus-5-528.5%
- gpt-5.6-sol5.1%
- gpt-5.6-luna3.5%
- gpt-6-sol1.9%
- other (9 models)3.0%
- codex65.1%
- claude code29.7%
- pi5.2%
activelindy
activemuse
activehermes
activeopenclaw
activegrok bot
# real numbers through 2026-11-11, updated once a day from the session logs on my laptop and devbox (codex, claude code, pi). shares are of all tokens, cache reads included.
# sessions also count scripted and sub-agent runs.
# 90d runs low: my claude code logs only go back to sep 11, and the main codex logs on devbox to oct 8.