Agent workflows
My daily agent stack (and what it costs)
Everyone writes about which model is smartest. Almost nobody writes about what the actual day-to-day stack looks like. Here's mine — it runs this very blog — with the honest parts included.
The three agents
- Desktop coding agent on my workstation: owns the site repo, does the HTML/CSS, runs the deploy script, holds the SSH keys. This is the "hands."
- Server agent on the VPS (CLI over SSH, inside tmux): watches logs, checks nginx, does server-side chores. It reads/writes our shared coordination folder via rclone. This is the "ground crew."
- A consumer AI assistant for research and drafting: no repo access, no shell. This is the "intern with a phone."
They coordinate through a plain Google Drive folder of timestamped markdown messages. No fancy orchestration framework — just files, naming conventions, and lane discipline. It's dull, and that's why it works. The failed experiments were all versions of "let the agents talk to each other directly"; every one of them ended in a permissioning rabbit hole or a loop.
The honest parts
- The cheapest model wins most days. Given the current price spread, the flagship models earn their cost maybe 5% of the time — hard refactors, gnarly debugging. Routing by task, not by brand, cut my effective spend dramatically.
- Lane discipline is the whole security model. The server agent can't deploy the site; the desktop agent doesn't touch server config. One agent, one blast radius. The 212-page system cards labs now publish are aimed at model risk, but most real-world risk lives in agent permissions, which is on you.
- File-based coordination beats APIs for this size. Message-per-file means diffs, history, and humans can read it too. When this breaks (it will, at higher volume), the fix is probably a queue — not an "agent framework."
Total monthly cost
Under $5 in API spend for everything that runs this site, plus consumer subscriptions I'd pay for anyway. The expensive part is not tokens — it's the hour a day I spend writing good task descriptions. The agents didn't remove the work; they moved it up a level.