The dashboard my agents built for themselves

I run a team of six AI agents on a Discord server. They help with my job hunt, my portfolio, my content projects. For months, everything they produced was scattered across chat threads and folders, and finding anything meant scrolling through conversations I'd half forgotten. So I did what any manager would do. I assigned the problem to two of them.

I asked my coder bot and my chief of staff to design and build a private dashboard that catalogs all of their work: projects, writing, applications, experiments. I set the direction, reviewed the pull requests, and tested the builds. They did the building. The result is Hermes Hub, a local-first Next.js app backed by SQLite, running on loopback on my own machine. Nothing leaves the computer.

Today I'm publishing the full case study: The office my agents built for themselves. What makes this write-up different from a launch announcement is the standard it holds itself to. "An agent built it" invites skepticism, and honestly it should. So I verified every claim the way I'd want someone to verify mine: 715 passing assertions across 62 files. Zod-validated inputs. Merges through pull requests with independent review. Each evidence claim walked back to the commit history, the PRs, and the validation docs.

It's also honest about what broke. My favorite section to write was the one about the stale Projects page, a caching bug where the page served old data after every restart. The fix was small. The lesson was bigger: agent-written code fails in the same boring ways human-written code fails, and it deserves the same boring rigor.

Here's what I keep coming back to. Directing agents turned out to be a management skill, not a prompting skill. The agents that did the best work were the ones with the clearest briefs, the tightest review loops, and explicit permission to push back when my direction was wrong. That's not an AI insight. That's just management.

If you're building with agents, or managing people who are, the case study is my attempt to show the receipts instead of just the demo. A narrated video walkthrough is coming next. Read it here: https://work.nikema.dev/case-studies/hermes-hub