Your Agents & What They Won't Do

Arc runs on eight named agents. Naming them is not decoration — a named agent can be described, configured, held to a limit, and switched off. "The system" can be none of those things, and you cannot tell a client that the system will not reject their candidate.

You can see all of them under Your agents in the sidebar.

The eight

They share one record — the mandate, and everything that has happened to each candidate. That shared record is what makes them a pipeline rather than eight separate tools: when Pulsar hands a candidate to Lyra, nothing is lost in the handover.

What is true of every agent

No agent can reject anyone. No score and no model inference disqualifies a candidate on its own. A person decides, and the database refuses the write without their name on it.

This is not a policy we intend to follow. It is enforced in the database, and a second check runs on every code change. When an agent forms the view that a candidate should be rejected, it records a suggestion and stops.

Each agent's own page ends with what it will not do. That section is the one you can repeat to a nervous client.

Needs a person

Two things only a human can settle appear in one queue, under Needs a person:

1. A candidate asked to speak to someone. They replied HUMAN, or used "I need help" during their interview.
2. An agent proposed a rejection. You see the reason it gave, and you decide. Rejecting requires you to type why, and it is recorded against your name.

The list is ordered by how long each item has been waiting, longest first, with a candidate who is waiting on a person ranked above a decision that is waiting on you.

A candidate who asks for a person is routed to you in every setting, including when the offer is not advertised anywhere. The setting controls whether we tell candidates they can ask — never whether we honour it when they do.

What Nova decided

Nova reviews every active role overnight and decides one next thing: source more people, enrich the ones you have, start outreach, try another channel, or hold.

You can see those decisions in three places:

• On the job's Overview tab, under the role — what Nova decided for this role, and how many of them are waiting on you.
• On your dashboard, when something is waiting: "Nova is waiting on you".
• In the daily digest, as a line in "While you were out".

Two things worth knowing about how this reads:

On Safe, Nova recommends and never spends. A list of recommendations that have not happened is the switch working, not a stuck pipeline. Acting on them is your call. Switch the role to Autonomous and Nova acts within the limits you set.

A repeated recommendation is counted, not repeated. Nova re-decides from scratch each night, so a role in a stable state produces the same recommendation every night. Rather than showing you the same line seven times, Arc shows it once and tells you how long it has gone unanswered — "try another channel, ×5, nightly since 25 Jul". That count is the useful part: it is the difference between a fresh suggestion and a question you have been ignoring for a week.

The decision card

At the top of every candidate profile is a single card: what we know, what it adds up to, and what is still missing.

• It recommends. It never decides, and it has no button that changes anyone's status.
• Missing is shown as missing. "No evaluation yet" and "scored 0 out of 10" are opposite facts, and the card will not confuse them. The same goes for a candidate nobody has judged against your ideal profile yet.
• It shows how much of the evidence we hold — because "this candidate looks weak" and "we barely know anything about this candidate" call for opposite actions. Below halfway, the card declines to judge in either direction and says so.
• Your own on-profile judgement sits beside the match score. When the two disagree, that disagreement is the most useful thing on the page.

Below the card, the Activity tab shows everything that happened to the candidate and who did it — by name, agent or person.

A score only covers what was asked

Every voice screen is scored on six things. If the conversation never got to one of them, that one is shown as not asked — an empty bar, no number.

This sounds obvious and it is easy to get wrong. A dimension nobody raised used to come back as a low score, and on the profile a 20 that means "we never asked" looked exactly like a 20 that means "we asked and the answer was poor". They call for opposite actions.

So Kepler is allowed to say it does not know:

• The overall score is averaged over the areas that were actually covered, not over six numbers where three were guesses.
• A screen that covered too little to judge is flagged as thin rather than scored low. The candidate is not marked down for a short conversation.
• On the comparison view, a candidate with an unasked area does not lose that row to someone who was asked.

The interview itself now puts one anchor question to each of the six, and pushes past the first answer on the four where a single reply settles nothing.

What candidates see

Before any interview, candidates go through four short screens: what the role is and how long it will take, their details and consent, three things worth knowing, and a ready screen.

"I need help" is on every one of them, and stays available during the conversation.

Silence is not a bad interview

There used to be a sound check in the setup steps. It was removed — it was the only screen that asked your candidate to perform something, and it sat exactly where a flow loses people.

The problem it existed for is real, and is now handled where it belongs. A muted hardware switch grants microphone permission and then sends digital silence, so the candidate says nothing while the AI keeps asking. Left alone that produces a completed interview with an empty transcript, which gets scored, and the candidate comes out looking weak.

Arc now checks that the candidate actually spoke before it will treat an interview as finished. Silence is recorded as interrupted, never scored, and does not use up their attempt — the same treatment as a dropped connection. The protection moved from your candidate's shoulders to our server, where it costs them nothing.

If their connection drops

An interview that is cut short is recorded but never scored, and it does not use up the candidate's one attempt. They can simply rejoin.

This matters because the alternative is silent and unfair. A candidate whose mobile signal dies two questions in would otherwise get a completed interview, a score computed from two answers, and no way back — and on their profile that reads as a weak candidate rather than as a bad connection. Scoring two answers out of ten is not a partial judgement; it is a made-up one.

The transcript is kept, because it is evidence of what happened.

If they opened your link inside another app (WhatsApp, LinkedIn, Instagram), Arc tells them to open it in a real browser before the interview rather than after it has failed.

Consent

One sentence, ticked before the interview starts:

I understand this conversation is recorded to assess my suitability for this job.

Video swaps one word. Arc stores the exact wording each candidate saw, so if you change the text later, older consents remain attributable to what those people actually read. Agencies with their own approved wording can supply it.

Interview prep

For the interview you run, Arc builds a prep plan: what the role asks for that nothing in the file answers, and questions written against exactly those gaps.

The gaps are worked out from your own records first — the mandate's must-haves, the job's requirements and responsibilities, everything on the candidate's profile, and the screening transcript — and only then are questions written for them. It is not a generic competency list, and a question that does not close a real gap is discarded rather than shown.

If the job lists no requirements, Arc says so rather than reporting that everything is covered. Those are opposite facts: one means the candidate looks strong, the other means we checked nothing. When it happens, the fix is to fill in what the role asks for on the job and generate the plan again.

Each question comes with what a strong answer contains and what a thin one sounds like. The page prints, for the room.

Open it from Interview prep on any candidate.

What candidates said

Under Insights, you can see how candidates rated the experience — how comfortable they felt, and whether they would recommend it — with the number of responses beside every average, because an average from two people is not a finding.

Comments are only shown if the candidate ticked the box saying you may share them, and are attributed by first name only.

Interview integrity

For unattended interviews, Arc records a few things about how the interview was taken: whether the tab lost focus for a while, whether fullscreen was exited, whether a long block of text was pasted.

This is deliberately conservative:

• One thing happening repeatedly is never enough. Ten tab switches is still the lowest level — a laptop that keeps sleeping is not a cheat.
• Ordinary behaviour is filtered out first. Glancing at the job description is not answer-shopping.
• The result is Clear, Low, Medium or High with the evidence listed — never a pass or a fail.
• It never rejects anyone, and it never changes the recommendation on the decision card. It appears as one line of context, and nothing at all when there is nothing to say.

Arc records that a long paste happened. It never records what was pasted.