nudge build log

  • 7th August 2026

My Mac taught my new Linux box my dev setup — one agent on each, talking over a relay

A case study in cross-machine peer agents: one agent on my Mac that knew my years-old dev setup, one on a fresh Arch box that knew Linux, connected over the same encrypted relay my phone uses. Neither could do the job alone. Together they didn't copy my config — they translated it, zsh to fish, brew to pacman, and caught bugs in the original on the way.

Read more 
  • 23rd July 2026

I didn't build an auto mode — I used a supervisor instead

Auto mode in today's coding agents means vacating the reviewer seat. A case study in doing the opposite: the agent I was already talking to stayed in the reviewer seat as supervisor while a spawned worker did the work — the supervisor approved 170 tool calls, ruled on the hard decisions, and escalated only what a machine can't know.

Read more 
  • 10th July 2026

An agent is just another client

How a small bug in my hand-rolled coding agent turned into one idea: there should be exactly one way to talk to an agent, and it shouldn't matter whether a human or another agent is on the other end.

Read more 
  • 5th July 2026

I made a coding agent you can drive from my phone

A build story: how decoupling the agent loop from the terminal turned into a session you can scan onto your phone and drive over an end-to-end-encrypted link.

Read more