Where I'm challenged, and what's next

The gaps I know about, and the order I plan to close them.

1Challenges

Real-environment validation. Several of my tools have run only offline, on mock data or in a lab. The AI zone is not committed, and the node script's generic-PC profile has never met real hardware. These wait on lab time, not code.
Outcome metrics. Almost every number on this site is an input: tests, reviews, findings. I don't have an hours-saved or defects-avoided figure yet.
Agent security backlog. Close the tracked hardening items, finish the read-only AI logins on the firewall, then commit and test the AI zone.
Doc drift. Agents write status notes faster than reality changes. It took two multi-agent truing-up passes to bring my docs back in line, and agent-written status reports once said "running" when nothing ran.
Usage limits. Frontier-model usage windows shape what can run, and multi-agent workflows have died mid-run on limits. That is one reason a delegation tool now sends bulk reading and first drafts to local models.
Over-building. I have built scaffolds and governance layers that I later retired. The lesson: audit what is already there before building more.
Owning what the agents wrote. Most of the code is AI-written. Before any of these tools touches a real system, I want to be able to walk a reviewer through its critical paths myself.