Nothing goes out
without my yes.
AI agents read my job mail, find the roles that fit, pre-fill applications and stop before Submit, and draft my replies and briefs. A Fact Checker checks every draft, Warden governs every agent, and nothing goes out without my yes.
- Agents
- 10, each governed in Warden
- The rule
- Nothing goes out without my yes
- Status
- My workspace · public sample on made-up data
Today: only what needs me.
Why I built it.
I wasn't tracking my job search at all. I applied to roles that fit by talking each one through with AI, one conversation at a time. After I built Warden and started applying seriously, I was more optimistic about it. Then I started using Clera, and it hit me.
The moment it started:
Why not build a job search dashboard and automate all the boring parts I hate to do? I can still check and approve everything, but I don't have to do any of it by hand.Me
So I started one agent at a time, and the agents built up from there.
An agent can draft.
It can't speak for me.
I gave every agent a job and a line it never crosses. The Inbox Watcher reads job mail and never writes or sends it. Job Scout finds and scores roles and never applies or contacts anyone. The Pre-filler fills the form in my Chrome and never submits, signs in or creates an account. Interview Prep never invents a fact: anything its sources don't back becomes a question in the brief. The Follow-up Writer sends only what I approved. The Clera Answerer drafts my answers to Clera's questions and never submits one without my yes. The AIApply Monitor tracks what AIApply sent and what came back; it only reads, and never applies or sends. The LinkedIn Agent handles LinkedIn outreach and never sends without my yes.
Each agent, and what it may never do.
Stop before Submit.
The Pre-filler does the slow part of an application: it tailors the résumé, fills the form in my Chrome and uploads it. Then it stops. The last click is mine.
Pre-filled, checked, and waiting for me.
A fact checker on everything
written in my name.
Drafts, briefs, tailored résumés and pre-filled answers are checked against one facts file and their cited sources before I see them, and anything that doesn't hold is flagged in amber. A second agent compares the agents' records with each other and with Warden's audit, and each disagreement comes to me as a Yes or No.
Warden decides,
and a send waits.
Each step asks Warden first. Reads still run if Warden can't be reached, and a paused agent stops. A send is different: my press of Send records an approval in Warden, the sender carries it, and when Warden can't be reached, the send waits instead of going.
From first commit
to a public sample in a week.
First commit on October 9, 2026; public sample live on October 15. I set the architecture and the behavior, the agents wrote the code, and I tested and deployed it.