Twelve named agents, and where each one really stands

An agent here is a bounded piece of work with a defined input, a defined output and a place in the record where its output lands. The Match Scorer takes a candidate and a requisition and produces a scored breakdown. That is the whole contract. It does not roam, it does not decide, and it does not act on the pipeline on its own.

The word is used elsewhere to mean an autonomous system that pursues a goal across your data and your outbound channels without a person in the loop. We are not building that for recruitment, and not because it is hard. A system that can independently contact candidates, move them through stages and dispose of them is a system nobody can reconstruct afterwards, in a domain where reconstructing a decision is regularly a legal requirement. Agents here propose. People dispose, under their own name, with a coded reason.

This is the full agent set as designed, with the status each one carries today. Live means it is in the product now. In build means active development for the Wave 1 release. Planned means designed and scheduled, with nothing shipping yet.

The ordering follows the desk rather than the roadmap, from intake through to coaching, because that is how the set is meant to be read: as one continuous workflow with several of its steps still missing.

Job Intake, Outreach Personaliser, Reply Handler and Interview Scheduler are being finished for Wave 1 because each of them raises the quality of something that already ships. Structured intake is what makes matching worth trusting, since a requisition written as prose can only ever be matched loosely. Personalised outreach and reply classification are what turn a shortlist into conversations. Scheduling is where a booked interview stops being an email thread.

The agents left for later waves are the ones that depend on data the product has not accumulated yet. Placement risk, client health and recruiter coaching are all pattern detection over your own history, and a model that raises an at-risk offer after two placements is guessing. They are worth building once the record has enough in it to be worth reading.

Every agent output lands somewhere a person can see it before it counts. A match score sits on the candidate. A shortlist proposal waits for an approver. A voice interview produces a transcript that a recruiter reads. A drafted message is a draft until somebody sends it.

The audit trail is the other half. Every stage change records the actor, and an agent is never the actor on a disposition. Where a run happened, the record shows the configuration in force, the set considered and the person who approved the result. That is what makes an agentic system reviewable rather than merely fast, and it is why the shortlisting agent ranks and never rejects.

Agent work consumes AI credits at a published operation weight, and the weight table is visible rather than described as usage. Professional includes one hundred credits per user per month; further credits are an add-on with a published price rather than a quote. Any operation that will consume a meaningful number of credits shows the cost before it runs, which matters most on batch work like importing an archive or scoring a large candidate set.

Every agent can be turned off. That is a tenant setting, a per-client setting and a per-requisition setting, because the requirement usually arrives from one client rather than across the business. When an agent is off for a requisition, the requisition records that it was off for the whole time it was open, which is frequently the thing you need to be able to show.

Live means the workflow is in the product and can be used today. Shipping and being extended means it works and is being widened for Wave 1. In build means active development against the Wave 1 release, written in the future tense everywhere on this site. Planned means designed and scheduled, with no code you can use.

Every feature page carries its own badge and, where the status is anything other than planned, a line naming the files or functions the capability lives in. If you find a page on this site claiming a capability without that evidence line, tell us, because it is a bug in how we write rather than an oversight we are comfortable with.