I stopped pitching agent magic and started writing contracts

I stopped pitching agent magic and started writing contracts

The first version of OpsPilot looked good in a demo, then failed consistency checks on repeated tasks. That is when the project moved from presentation mode to contracts-first engineering.

I built a local-first workflow loop around observe, propose, execute in allowlisted paths, and write-back logs. YAML packs define flows and SQL memory tracks state across runs.

The biggest lesson is that convenience shortcuts become incidents. The useful work is explicit success and failure contracts per step, with tests around risky IO edges.

Current status is a working prototype with a hardening list. Next step is more packs, tighter failure mode handling, and cleaner run diffs. GitHub: https://github.com/pradyotbathuri