engineering2 min de lectura
Friday Five #3: The hands are replaceable
How I route work between coding agents, the gate that blocks deleted tests, and why Claude Duet and OpenAI Dots come down to the token.
Rob Vasquez·
Five notes from the week. Short on purpose.
- Using: One fixed planner, swappable coding agents. Each kind of task has a primary host and two fallbacks. I don't trust vendor labels anymore. A model listed as certified for agent work answered five identical probes with 200, 200, 429, 429 and 404.
- Using: A push gate blocks any commit that removes, skips or comments out a test unless I write a one-line reason. I replayed it on my last 100 merged PRs and it would've stopped five. Agents don't get to write the reason.
- Reading: Why I keep agent notes in two places. Memory holds credentials, behavior rules and my preferences. Retrieval holds anything about the code, because it stamps every hit with an age and memory doesn't.
- Reading: Claude Duet (Oct 21) and OpenAI's Dots (Sep 29). Same bet, an agent on its own cloud computer, and the same harness question: who approves, and what can the token reach? Duet waits for your OK before it sends, pays, signs or deletes. I built a read-only Dots connector first, and it never went live. The approval screen is the promise. The token audience is what keeps it. Full comparison on robslog.
- Asking: If your main coding agent went down tomorrow, which kind of task wouldn't have a second host?