Screen a big pile of job candidates with one agent per person
Idea + starter prompt, not run by us yet
For: Teams that get more applicants than they can read and want a fair, checkable shortlist
A team exported its hiring database to one file per candidate and ran an agent on each of 115 people at once. Each agent scored its person against written rubrics and checked the GitHub and portfolio links. Code did the sorting, extra agents argued against the highest-scoring few, and the whole run took about 13 minutes at a modest cost.
Steps
- Export your applicant list to one file per candidate, and write your rubric as plain markdown files.
- Do a small test with 3 to 5 candidates first and read each score and the reasons given.
- Have each agent return structured scores only, and let a script sort people by how many rubric points each meets, not the model.
- Add a second pass where separate agents try to find holes in the highest few scores.
- Cap how many people can reach the highest grade, then make the hiring decision yourself and read the shortlist before anyone is contacted.
- Watch the bill. Parallel agents each build their own cache, so costs add up. The authors suggest a cheaper model for scoring and the stronger one only for the review pass.
Starter prompt
Trial first. In a new folder called [trial], make 3 made-up candidate files and run the whole job on those only. Save the trial results in [trial-output], never next to the real files. Work on copies of the real files from then on. My candidate files are in [candidates folder] and my rubrics are in [rubric folder]. Show me your plan first: how you score each person, the output format and a rough cost estimate. Wait for my OK. For each candidate, read the rubrics, check any public links in the file and return JSON with a score per rubric item plus the evidence you used. Say plainly when you could not verify something. Then sort people by how many rubric points each one meets. Use a script for the sorting, not your own judgment. I make every hiring decision. Do not contact any candidate and do not change the originals. Stop and wait for my yes before anything is sent to anyone. Showing me a list is not permission to go ahead.
Original post: happycapy.ai · found via agentusecases.ai. Written up in our own words. We haven't run it, so test it on made-up data first.