The ATS Agent

What it is, how it works, and how your team talks to it.

What it is, in one paragraph

The ATS Agent is a standing recruiting agent, built on the same framework as Pam (your CS agent). It works two jobs at once: an ATS director that advises you on hiring decisions (funnel metrics, req priorities, offers, compliance), and a recruiting manager/interviewer that helps the team execute (screening, interview prep, candidate comms). It reports to Toby, the same way a recruiting coordinator would report to your HR manager. It has its own memory, learns from feedback your team gives it, and every rule it learns is proven before it ships, not just accepted on faith.

How your team talks to it

When Jose or Homer ask Toby a recruiting question in Slack, Toby does not answer from his own memory. He is required to hand the question to the ATS Agent, get its answer, and relay it back like a manager would. This was a deliberate build choice: Toby already talks to every employee with no restrictions, and the Vegas AM search running through the ATS Agent right now is confidential. Piping the question through a locked door (see below) means Toby can serve everyone else normally while this one topic stays fenced.

Where you and Jose reach it directly

The shared ATS page (Vegas Hiring tab) is where you and the panel work with it hands-on:

How it actually learns (not just "we told it something")

A note in the teach box does not become policy just because someone approved it. Before a taught rule ships to the agent's live memory, it has to answer 2-3 real past cases it was never shown, and an independent grader checks its answers against what actually happened. Pass, and the rule ships tagged with proof (like learned 3/3). Fail, and it comes back to the Today page with the case it missed, so you can see exactly where the lesson didn't stick. This is the difference between "I taught it" and "it generalized," and it is the part of the build most likely to quietly break if anyone ever touches the code without re-running the proof.

What changed in yesterday's hardening pass

A full two-model code review (two different AI model families, so blind spots do not overlap) found 8 real problems in the build. All are fixed and re-verified live:

• Fixed Homer could previously approve rules straight into the agent's authoritative memory, more access than your own team had through Toby. Now locked to you and Jose.
• Fixed A channel-thread Slack conversation could let someone else's reply ride an authorized person's identity into the confidential search. Delegation now only works in a direct message.
• Fixed An approved rule's text was never actually run through the compliance checker before shipping. It is now, and it fails closed (holds the rule) if the checker can't run.
• Fixed The spend safety cap (meant to halt the agent if it runs too often in a day) was silently broken, watching a log file that never existed. It now counts real runs correctly.
• Fixed Two people clicking Approve around the same time could silently lose one click. Both write paths now share a lock.
• Fixed today The compliance checker's age-question detector only caught one exact phrase. Widened to catch birthdate, graduation year, retirement-year questions, and a few other protected-class categories that were similarly narrow.
• Fixed today The ATS page's buttons are now part of your nightly automated build-verify check, so a broken button gets caught overnight instead of by someone clicking it.
• Fixed today When Toby relays a candidate-facing draft (an email to a candidate, for example), it now also passes a stricter candidate-communications check, not just the internal one.

Where things live (for whoever touches this next)

Still outstanding

Smaller, optional items: the Chargeback Agent (a separate, unrelated build) has a cron that never got loaded, unrelated to anything here. And two open questions from the original build review are still unanswered: whether flags should also ping Slack, and whether the agent should triage the Vegas hiring inbox daily.