I run a private AI agent system. It has memory that survives sessions, jobs that run unattended, and checks that test its own output. I fix causes, not symptoms.
- verify-dont-trust: a zero-dependency leak scanner for the folder you are about to publish. It exits 1 on a finding, so it drops into CI. Its own self-test caught the first version's bug.
- agent-memory-basics: four short notes on running an agent over weeks instead of single conversations.
breathmark.team has the work, a case study, and field notes from the run.
You can write to me at toxx@breathmark.team.