Pete
15 August 2026 · Agent system · 1 min read
tl;dr
Four agents that take one job application from posting to scored resume, and hand every judgement call back to me.
- Pulls the exact terms a screening system searches on
- Finds who the role reports to before I write to anyone
- Tailors the resume without claiming anything I cannot back
- Scores the result and names what it could not cover
keyword finder
Reads the posting and pulls the exact terms a screener searches on
manager finder
Works out who the role reports to and writes the searches to find them
resume optimizer
Rewrites the resume against those terms, claiming nothing I cannot back
scorer
Simulates what a screener extracts, then names what is missing
01The problem
Applying properly to one job takes an evening. Reading the posting closely, working out what it is really asking for, finding who owns the role, rewriting the resume, then checking whether that rewrite survives the screening software. Doing it badly is fast. Doing it well does not scale.
02What I built
Four agents, each owning one part of that evening, running in order with me in the middle of them.
03How it works
Each agent gets one job and a narrow set of tools, which is what keeps them honest. The optimizer sorts every keyword into what I can actually claim before it writes a word, so it never invents experience I do not have. The scorer then opens the finished file and simulates what a screening system would pull out of it, which is a different question from how the page looks to a person.
04What I got wrong
The first version ran straight through without stopping. That was the mistake. Which role to target and which experience to lead with are judgement calls, and automating a judgement call only means making it badly at speed. Now it stops and asks.
Thoughts on this one?
Send it straight to me. Nothing is published here, it just lands in my inbox already labelled with which build you are talking about.
Comment on Pete