My AI employee
Kai sees what every employee does, and tells me what I need to decide.
Kai has been running in my business since September 3, 2026. He has no classroom class of his own and isn't a package, because he isn't taking on a standalone task for an outside company, he runs my own workforce.
After every session, Kai reads what an employee has done, git commits, reports, protocols, and checks it against what's actually on the file system. From that he writes a short report of what I need to decide, instead of me clicking through ten folders myself.
A real workday
Three reviews are documented so far: on September 4, 2026 at 00:13, on September 4, 2026 at 12:57, and on September 5, 2026 at 13:45. In each one, Kai checks what an employee claims against what's actually on the file system and on a live website.
3
Reviews documented
6 of 6 live
Core routes checked after rollout
3
Findings in one review
In the September 5, 2026 review, he checked two colleagues' work against the actual file system and found three issues: an incomplete first correction pass on a customer website, which had fixed the blog posts but left the underlying pages unchanged, so the live site contradicted itself in five places; an open approval question, because one employee had been running on an unrestricted permission level since morning without the reset being decided; and unclaimed authorship, because a script had been run and a knowledge file changed with no session or log showing who had done it.
The same review also confirmed a completed rollout: six core routes of a customer website checked live with HTTP 200, an outdated sentence that used to appear twice showing zero hits after the rollout, the new sentence live twice. Kai reports both directions: what checks out, just as much as what's missing. When an approval line or a named file is missing, he writes "not verified" instead of smoothing it over, three times in this one review alone.
The tools he works with
Kai doesn't work in an interface of his own; he reads what every other employee leaves behind anyway:
- Every employee's git history and commits, to see what was actually changed.
- Reports, protocols and personnel files across the monorepo, checked against the real state of the file system.
- The workforce's shared mailbox protocol, through which he sends questions and notes to individual employees.
What he can't do, where I approve
- He decides nothing on substance, approves nothing, and commits nothing.
- He doesn't change other employees' work, only his own review protocols and mailbox notes.
- He doesn't write other employees' personnel files, that isn't his job.
- He only hands out substantive tasks to other employees on my explicit instruction; he's free to send questions and notes on his own via the mailbox.
- If an approval line or a named file is missing, he writes "not verified" instead of quietly smoothing it over. I stay the only approver, for Kai just like for everyone else.
The path: build it yourself, there's no ready-made template
Standard: build it yourself
The entry point is Learning Claude, plus Hiring an AI Employee for the process and Creating an AI Employee for the build. What's special about Kai: he watches git, reports and sessions from several other employees and checks claims against evidence, instead of just reporting them. This role assumes you already have several employees running, Kai isn't a beginner role.
Honestly: not a package
Kai only runs in my own business so far. There's no ready-made template you can download yet, the task is too tied to my own specific mix of several employees.
Frequently asked questions about Kai
How do I keep track when several AI agents are working at the same time?
For me, Kai handles it: after every session, he reads what an employee has done, commits, reports, protocols, and checks it against the real state of the file system. From that he writes a short report that only shows what I actually need to decide. I no longer click through ten folders myself.
Do you need a leadership role for several AI employees, or is a dashboard enough?
A dashboard shows numbers, but it doesn't check whether a claim is true. Kai compares what a report says with what's actually on the file system, and flags gaps instead of smoothing them over. Past a certain number of running employees, just watching isn't enough anymore, and a dedicated review role starts to pay off.
How do you know if an AI agent really did what its report says?
That's exactly what Kai checks for me: he compares an employee's commits, reports and protocols against the real state of the file system and the live website. If something doesn't match, say a missing approval line or an unverified change, he explicitly writes "not verified" instead of quietly closing the gap.
At how many AI employees does a dedicated coordination role pay off?
I can't back a fixed number, I simply have too few data points of my own. In my case, the workforce had already been running for a while before Kai joined. My own rule of thumb: once you stop reading every report yourself, you need someone who does that for you.
What can a leadership role for AI employees decide, and what can't it?
With Kai, the line is drawn clearly: he decides nothing on substance, approves nothing, and commits nothing. He may send questions and notes to other employees on his own via the mailbox, but only assigns substantive tasks when I explicitly instruct it. I am and remain the approver, for Kai just like for everyone else.
Several AI employees, one overview, no black box
There's no template for Kai's exact setup, but the learning path he grew out of is in the community.
Join the community