|5 min read
Parallel Agents Need More Than Separate Branches
Separate worktrees help with file edits; an ownership map makes shared databases, ports, contracts and integration checks explicit.
[AI & Data][Software Engineering]
Writing about AI, infrastructure, and what actually ships.
total 160 · ~13.0h read · updated Nov 13
A release receipt should connect the tested candidate, built artifact, deployment and production route without pretending one smoke check proves everything.
A 93% approval rate can't tell a careful reviewer from a rubber stamp, so keep people at the plan and the signature, enforce the rest in configuration, and test the remaining gates with seeded known-bad cases.
A decision table for retrying agent actions when the response is lost, with stable intent, provider-specific idempotency and an explicit unknown outcome.
A Raspberry Pi Zero 2 W and a seven-colour e-ink panel turn my remaining Claude and Codex allowance into artwork: tentacle coils, speech balloons, reactor tanks and moth wings. ImageGen painted the scenes; code draws every live number.
I let coding agents research, source, generate and iterate three native KiCad revisions of an NFC business card over three days in October. The files are real and the checks are clean; the confidence came from datasheets, exact part numbers and independent re-examination, not from the renders.
I wrote a Claude Code skill in April because I book the same gym classes every week. By October it was an app on both stores, and an agent runs it. Here is why.
I ran Together's HybridAgent checkpoints with latent reasoning on and off across 72 requests in 36 on/off pairs on two 2016 GPUs. The visible token rate fell; correct answers sometimes arrived twice as fast.
A technical follow-up to the Reverse Tamagotchi: physiological days, missing scores and stale rows can change what a fitness assistant thinks it knows.
A measurement protocol for human review, rework and agent spend that keeps rejected attempts in the calculation.
A durable handoff separates observations, hypotheses and unfinished work, then records what would make each observation stale.
The $7.99 rebuild bought a first version; the repository now shows the less photogenic work of preserving an archive and keeping its promises.