Kesha SelectAI agent operations
I run the store's daily checks and growth reviews through AI agents that audit themselves
The store was run by hand, with no ads or search program and little data. I built agents that check, draft and report across the store, search, email and ads, plus a daily review that proposes fixes to the system itself.
Scope
Store to ads
Agents scan the storefront nightly and review search, AI-answer visibility, accessibility, email and ads.
Safety
Single writer
Only one agent can change the store, and every write is backed up and can be undone.
Self-improving
Daily audit
Each day the system reviews its own work and corrects itself. Major changes wait for my approval.
How I built it. Built on Claude Code: I designed the system, agents and approval rules; agents apply routine corrections themselves, and major changes wait for me.
- Work starts on its own: nightly site checks, plus scheduled search, email and ads reviews.
- Drafts arrive as proposals; spend, discounts, emails and customer-facing changes wait for approval.
- The dashed loop: the system correcting itself; major changes come to me.
What it caught
It flagged a nightly check that had stalled for 9 days and retired it, and moved 7 scheduled jobs to Pacific time before daylight saving ended.
The decision
Chose
Most agents can only read. One can change the store, and every change it makes is backed up and can be undone.
Over
Letting every agent write to the store.
Because
A single door is one place to audit and one place to undo.
Deferred
More agents already specified (growth, campaign planning, service operations), not built yet.
Team
Built solo, reporting to the CEO.
I owned the system design, the agents and the approval rules.