Building thoughtful software with AI.Notes from the systems behind the work.
A self-repairing deploy step grabbed my Proxmox host's root key whenever it saw drift, despite docs saying a human runs it. I swapped it for a token that can only run guest commands on one VM, pinned the API certificate before the token is ever sent, and banned the old path in the deploy scan.
Read moreThe code that copied an agent's sandbox work into a pull request only knew how to upload files, so anything the agent removed stayed on the branch. Now deletions get their own lookup and delete step, though renames and tests are still open.
Cutting a single-file HTML pitch deck into per-slide files with a one-time extractor and a small build script exposed slide numbers kept in four places, a string-replace dollar-sign trap the deck dodged by luck, and an extractor still able to overwrite edits.
A fake fix button became a real pull-request pipeline, and then I found the guardrails it generates fall back to a one-word placeholder, while the template tests cheer for that placeholder and fail a real detector.
A committed run log shows my guidance-writing agent producing reports about its own work while the pipeline logged nothing changed. Reading the diff, most of the decisions that matter take the agent's word: model-picked confidence, an unreadable answer scored as perfect, and file paths scraped from logs instead of the diff.