I spent more time fixing my AI employee last week than fixing my actual product with live users.
That's when my AI-native workflow got too meta.
I’ve been building Maibel while also building Dobby, my internal AI ops agent and cross-channel observability layer.
He watches user issues, runtime signals, ops threads, and manual workflow tests, then turns the mess into evidence I can act on.
Incredibly useful. Also, a second product to maintain.
Two weeks ago, he burned through nearly $1.2k in 12 days. Last week, the issue was founder runtime.
In 7 days, we moved 47 issues:
22 were Dobby/ops-related: cost control, context hygiene, canary routing, and observability.
25 were still core Maibel build: infra, behavior, payment, and memory.
The week didn’t disappear into Dobby maintenance. It exposed the real problem.
I could now see more of the company, but still hadn’t routed the work cleanly enough.
The bottleneck was me.
My workday: Codex on VS Code for Maibel's core build, Claude Code SSH’d into Dobby’s VM, marketing/distribution work running on the side.
Everything was technically AI-assisted.
Somehow, everything still queued behind me.
That was my wake-up call.
I’d been treating AI tools like extra hands, when what I needed was a control layer.
So I started routing my stack by risk, context, and permission level:
1. Product code needs a high-control lane because live users are on the other side. Coding agents speed up Maibel’s build but I stay close to tests, diffs, Git PRs, staging checks, and final release judgment.
2. Ops needs a monitoring lane. Dobby works across user issues, runtime context, Cloud Run, Supabase, and workflow testing.
3. Agent infrastructure needs its own sandboxed lane. Dobby stays isolated in his VM as my observability layer. A dedicated coding agent works on his core with no access to Maibel’s product code.
4. Marketing and research get a looser exploration lane. I’m exploring Hermes for distribution workflows, campaign angles, and pattern research.
5. Shared sources of truth matter most. My agents point back to Linear for live issues, Notion for the company wiki, and GitHub for implementation history.
My biggest risk now is getting lost across contexts.
My brain can hyperfocus beautifully, but twelve open workspaces can turn into a crime scene if the naming conventions are bad.
I’m still building the interoperability layer: agents entering the right workspace, reading the right context, and leaving evidence I can review.
AI-native work isn’t trying every new tool that drops.
It’s learning which parts of the company need tight control, which need operational observation, and which can tolerate exploration.
Some workflows shouldn’t go straight into product code. They should be tested manually, stabilised, then encoded once the pattern is clear.
My founding stack isn’t my tool list. It's my permission model.
My AI employee didn’t make me less responsible. He made the responsibility harder to hide.
Hi, I’m Mabel.
I’m the founder of Maibel. I write about emotional AI, women’s wellness and learning to build the things I wish existed.
More about me
