This page is the how-to for the AI Monitoring Agent. Read Agent cautions and limits before you raise its level, and Authorization levels and grants before you let it retry anything.
| You want to | Who can |
|---|---|
| Turn the agent on or off, choose its level, schedule and data, set quiet hours, grant or revoke authorizations, press Run now, delete agent data | An administrator |
| Approve or reject a proposed retry | An operator or an administrator |
| Read insights, give feedback, ask the agent, set your own preferences | Signed-in users of the organisation |
| Suggest a rule | Anyone. An administrator approves it |
Every plan includes the AI Monitoring Agent. Quotas, such as how often it sweeps and how long insights are kept, differ by plan, and the agent card shows yours.
One thing must be true before the agent can run: the organisation's AI features switch is on. It is one switch for the Copilot and the agent.
The card also shows when the last sweep ran and when the next is due.
Open Monitor and choose the Agent Insights tab.
You are told about new findings on the notification bell. A digest e-mail and a weekly summary are available if you opt in. The weekly summary is a count of what happened, not a list of errors.
Open a card to see the diagnosis, the evidence and the run behind it.
| Action | What it does |
|---|---|
| Acknowledge | Marks the insight as seen so it stops counting towards what needs attention |
| Snooze 24 hours | Hides it from the active list until tomorrow, then it comes back |
| Resolve | Marks it fixed. The agent also closes it by itself if the task later succeeds |
| Dismiss | Marks it as not worth acting on, and removes it from the active list |
Resolved and dismissed insights stay readable under the Status filter. You can add a note when you resolve or dismiss.
At the bottom of the drawer, Was this diagnosis helpful? takes a rating and an optional note. Rate honestly, including the misses. Ratings are how the agent's accuracy is measured, and they are never used to widen what it may do.
This applies at level 2 and above. When the agent decides a retry is worth trying, it does not run it. It appears under Needs your approval at the top of Agent Insights.
Approving retries the failed run once and then watches it. It is a real run on the platform and it counts towards your run allowance. Just before it runs, the agent re-reads the run and re-checks every limit, so a proposal can still be refused if things have changed since it was written. A proposal that nobody decides expires, and the next sweep can propose it again if it still makes sense. Your decision is recorded in the audit trail.
Ask the agent opens a chat panel on the dashboard and inside an insight drawer, with the insight you are looking at as context. Good questions are specific:
The assistant can read failed runs, run history, task configuration, dependencies, platform health and contention, and it answers from what it read. It can also propose a retry or a rule from the conversation. A proposal is a card that shows exactly what would happen, and nothing is written or run until you approve it. The conversation lives in your browser session only and is not stored.
Each person has a daily chat limit. When you reach it the panel says so, the rest of the dashboard keeps working and the limit resets the next day.
Rules and preferences on the dashboard opens what the agent has been told. A rule is an instruction with an author and a scope: the whole organisation, a platform, a job, a task, or one person.
The most useful rule is whether a retry is allowed. You can set it where you work, in the AI agent section of the Task, Job and Platform editors:
| Setting | Choices |
|---|---|
| Failed runs of this scope | Not set, the agent may propose retries, or never retry |
| Safe to repeat | Not set, yes (a timeout may be retried) or no (timeouts are never retried) |
Only an administrator can change them. Others see the current setting. A Never on a task, job, platform or the organisation always beats an Allowed anywhere else, and a person approving a retry cannot override it. See When is a retry safe? before you mark anything as safe to repeat.
On the agent card, Quiet hours sets a window in your organisation's time zone, every day, on weekdays or at the weekend, during which the agent will not retry anything by itself. Proposals still appear during quiet hours. Only automatic retries wait.
Use the gentlest control that does the job. Each stops more than the one before.
| To do this | Use |
|---|---|
| Stop automatic retries but keep the agent watching | Revoke the standing authorizations, or lower the level to 2 or 1, and save |
| Stop the agent completely | Switch it off on the agent card and save. New sweeps stop at once |
| Stop the Copilot and the agent together | Switch off AI features |
| Stop one run | Cancel the run. A cancelled run is never retried |
Polysync can also pause the agent for your organisation if you ask support. The card then says it is paused, and it does not sweep until it is resumed. If the agent has paused its own automatic retries, the card says why and offers Resume automatic retries. See Authorization levels and grants.
An administrator can choose Delete agent data on the card. It permanently removes insights, feedback, learned rules and failure statistics for the organisation. Run summaries and the audit trail are kept. See Agent data and privacy.
Check these in order:
If it still looks wrong, contact support at support@polysync.com.au and say which organisation, which task and roughly when the run failed.