AI Agents
Phase 5 replaces the old landing page with an Agent Dashboard — a set of AI agents that watch for work in your HR data, propose what to do about it, and wait for you to approve.
Each agent proposes. You decide. Nothing an agent suggests reaches an employee, a candidate or a policy until someone approves it.
Who can use them
Every agent, the approval Inbox and the agent history are restricted to the HR and Admin roles. Employees and Managers do not see the Agent Dashboard and cannot approve agent proposals.
The dashboard
The dashboard is titled Your agent workforce. Each agent appears as a card showing whether it is switched on:
| On the card | Meaning |
|---|---|
| Enabled chip | The agent is switched on and working |
| Open | Opens that agent's own page |
| Coming soon (greyed out) | Built but not yet available to switch on |
Agents are enabled individually, so you can start with one and add others later.
The three agents
Sourcing Agent
Watches your candidate pipeline, reads each candidate against the job description, and proposes the next move with its reasoning. It files four kinds of proposal:
| Proposal | What approving it does |
|---|---|
| Shortlist candidate | Moves a new applicant worth progressing to Shortlisted |
| Rejection reply | Declines a candidate still at Screening and sends the rejection |
| Interview invite | Moves a shortlisted candidate to Interviewing, applies the suggested slot and sends the invitation |
| Candidate re-engagement | Reopens contact with a candidate who went quiet |
Where a proposal sends something, approving it sends it — you do not go and send the email separately. An interview invite always carries a suggested time, which you can edit before approving.
Onboarding Agent
Watches the whole joiner-to-leaver path and chases what has stalled, so HR is not manually following people up. Six things it looks for:
| Proposal | When it fires |
|---|---|
| New joiner setup | A recent starter is missing mandatory policy assignments. Approving assigns them |
| Onboarding invite chase | Someone was invited but never activated their account |
| Onboarding checklist chase | A recent starter has activated but still has outstanding checklist items |
| Contract signature chase | A contract is waiting on the employee's signature |
| Policy acknowledgement reminder | An active policy assignment has gone unacknowledged |
| Offboarding completion | Someone's leave date has passed but their account is still active |
Every one of these is a chaser or an assignment. The agent does not set pay, and it does not set leave entitlements — those stay with you, on the employee's record.
Approving an Onboarding invite chase re-sends the invite email with a fresh onboarding link, valid for 14 days.
This agent acts on the same items staff complete under My Profile → Onboarding. See Onboarding Operations for what HR configures, and Onboarding for the employee's side.
Policy Agent
Reviews the policies you have already published and files what it finds as proposals in the approval Inbox. It reads only your own policy documents — see Policy Management for where those live.
It looks for five things:
| What it looks for | When it fires | What the proposal contains |
|---|---|---|
| Not landing | A mandatory, published policy still unacknowledged by at least half of an audience of 3 or more, 21 days or more after it was assigned | A clearer, shorter restatement of the passages most likely causing the confusion, drafted from the policy's own text |
| Stale | A published policy unchanged, or unreviewed, for 180 days or more | A short list of specific things to re-check — named laws, dollar figures, dated procedures or contact details that appear in the text |
| Coverage gap | One of six common policy types is missing: Code of Conduct, Leave, Workplace Health & Safety, Privacy & Data Handling, Anti-Discrimination & Harassment, IT & Acceptable Use | A generic starter policy. The proposal says plainly that it is not tailored to your company |
| Dormant | A published policy at least 14 days old with no employees assigned to it | A flag only. It never guesses who the audience should be |
| Policy finding | Reading your published policies together turns up a conflict between two of them, a gap against a common legal obligation, or an out-of-date reference | A short note naming the conflict, gap, or outdated detail, citing the passages it's based on |
Nothing is published on your behalf. Approving a coverage-gap proposal creates a new policy in Draft — never active, never mandatory. Approving a stale-policy proposal records only that a review happened. Approving a dormant flag records only that the gap was noticed. A policy finding is acknowledged or dismissed rather than approved — either way, nothing in your policies changes.
This agent reads your own documents and nothing else. It does not look up Australian legislation, the National Employment Standards, or any other external source. When a policy finding mentions a legal obligation or outdated legislation, that comes from the model's general knowledge and has not been checked against current law, and any citations point to passages in your own policies. Treat it as a reviewer that reads everything and never gets tired — not as legal advice.
The approval Inbox
Agent proposals collect in the Inbox. Each proposal shows what the agent found, what it wants to do, and its reasoning, with actions to approve or decline.
Working through the Inbox is the normal way to use the agents — you are reviewing a queue of decisions rather than hunting for work.
An agent's own page
Clicking Open on an agent card takes you to that agent's page, which is where its settings and its record live.
| Section | What it holds |
|---|---|
| Duty | What this agent is responsible for |
| Waiting on you | Its proposals that need a decision — the same items that appear in the Inbox |
| Recent activity | What the agent has done lately |
| Data it can see | The scope of HR data this agent reads. Worth checking before enabling an agent |
| Autonomy | Whether it must ask before acting — see below |
| Hera for your company (Hera only, in place of Duty) | Whether Hera is switched on for everyone in your company. HR admins get a Turn on Hera button here when it's off |
| Settings | The agent's own configuration |
Turn off stops the agent from that page.
Before you enable an agent, read Data it can see. It tells you which HR records that agent can read — the quickest way to satisfy yourself the access is appropriate before it starts working.
History and audit trail
Every agent keeps two records:
| Record | What it holds |
|---|---|
| History | What the agent did and when |
| Event trail | The decisions taken on its proposals |
Both are reached from the agent's own page, under Recent activity. Together they mean an agent's actions are always traceable back to who approved them — useful for compliance, and for understanding why an agent proposed something.
Autonomous mode
Each agent's page has an Autonomy setting, and it reads Always needs your approval.
That is the only setting available. Attempting to set anything else is refused with "Only 'approval' autonomy is available in this phase — every agent action needs a human decision."
So no agent can act without a person approving it, whatever the setting appears to offer. Autonomous operation is a future release, not something switched on today.
Hera
Hera has not gone away, and in Phase 5 it does more than answer. The message box prompt says so: "Ask a question, or tell it what to do…".
Hera splits what you ask into two kinds of request:
| You ask it to | What happens |
|---|---|
| Read something — a leave balance, a policy, an employee detail | It answers straight away in the chat, from your real HR data |
| Change something — approve a leave request, add a candidate, chase an onboarding item, sync to Xero | It does not do it. It files a proposal into the same approval Inbox the agents use, for an HR admin to approve |
So Hera never writes to your HR data on its own. A request to change something becomes a proposal like any other, with the same approval step and the same audit trail.
Employees can use Hera, but it only ever reads their own records. Anything wider is refused. The organisation a request belongs to is resolved on the server from who is signed in — never from anything the AI supplies.
What powers all this
All agents run on Xagent. Phase 5 also moved the AI features that already existed onto Xagent:
| Feature | Where you use it |
|---|---|
| Job description creation | Recruitment Management |
| Candidate reviewer | Candidates Management |
| Policy summarizer | Policy Management |
Those features work as before — the change is the engine behind them, not how you use them.
- What's New: Phase 5 — the full release summary
- Dashboard — the dashboard the Agent Dashboard replaces
- Roles and Permissions — the HR and Admin roles that gate agent access