The AI Monitoring Agent

The AI Monitoring Agent watches your task runs on a schedule. When a run fails, it reads the run record and the platform's error message, compares the failure with earlier ones in your organisation, and writes an insight that explains what it thinks went wrong and what evidence it used. Its findings appear under Monitor, Agent Insights, and optionally on the notification bell.

The agent is off by default. Nothing runs until an administrator of your organisation enables it and records consent to scheduled AI processing.

What it can do

Capability	What you get
Find problems	Failed runs, runs stuck in Running or Queued, schedules that should have fired and did not, several tasks failing together on one platform, tasks that fail and recover in turn, a failure straight after a run of successes, runs far slower than usual, and a falling success rate
Diagnose	The kind of failure (credential, permission, configuration, connectivity, capacity, data or transient), a plain-language explanation and the evidence behind it
Remember	Whether the same failure has happened before in your organisation, and what happened then
Tell you	An insight card on the Monitor page, an in-app notification, and, if you opt in, an e-mail digest and a weekly summary
Recommend	At level 2, a proposed retry that a person approves or rejects, with the reasons and the checks it passed
Act within limits	At level 3, when it is turned on, a retry under a standing authorization that an administrator scoped, limited and can revoke. Try it with caution
Follow up	After a retry it watches the run. If the run fails the same way, the agent stops and tells you instead of trying again
Answer	Ask the agent answers questions about failed runs, history, configuration, dependencies, platform health and contention, from the data it reads
Learn from you	Your rules, preferences and ratings shape what it says and what it may propose. It can suggest rules from outcomes, and a suggestion does nothing until a person approves it
Extra platform detail	For platforms that support it, more error detail fetched from the platform after a failure. This is optional and off unless you switch it on

Every insight card shows where it came from: AI-generated, Rule-based or Detector. A rule-based or detector card used no AI model.

How a sweep works

Read. The agent reads the runs that failed or look unusual since its last sweep, together with the context listed under What it reads, below.
Classify and detect. It groups the failure with earlier ones and checks for stuck, blocked, silent and slow runs.
Explain. It writes the diagnosis and the evidence it used.
Tell you. It opens or updates an insight, and notifies you. At level 2 and above it may also propose a retry.

The four authorization levels

Level	What it may do	Available
0 Off	Nothing. No sweeps are claimed, no model calls are made and no AI data is written	Always
1 Observe	Reads run history and failure data, diagnoses, writes insights and notifications. Changes nothing	Always
2 Recommend	Proposes a retry for a person to approve	Available when turned on
3 Act within limits	Retries by itself under a standing authorization an administrator has scoped and can revoke	Available when turned on. Try it with caution

Every plan includes the agent. Quotas, such as how often it sweeps and how long insights are kept, differ by plan. We recommend working up the levels: start at Observe, move to Recommend until its proposals match what you would approve, and only then try Act within limits, with a narrow, short authorization.

The only action the agent can take on your runs, at any level, is to retry a failed run. Every retry is decided by code against your rules, its limits and a set of checks, not by the AI model, and every one is recorded in the audit trail. See Authorization levels and grants.

What it reads

Run summaries for failed runs: task and platform names, statuses, timings, retry history.
Error messages, after known secret patterns have been removed. An administrator can switch error messages off for the agent entirely.
Task and platform configuration metadata: names, settings and attributes, never parameter values and never resolved secret values.
Platform health and contention state.
Statistics from earlier similar failures, and Polysync's own product documentation.

It never sends parameter values, secret vault values or anything identifying a person. The full list, and where each item is stored, is on Agent data and privacy.

What it cannot do

At Observe it cannot change, cancel, retry or delete anything, and it cannot change its own settings. At Recommend nothing happens without a person approving that specific action. At Act within limits only actions inside a recorded, scoped, expiring authorization are possible. It never edits a task, job, platform, schedule, credential or pipeline.

In chat it can propose a change, such as a new rule or a retry, but it cannot make one. A proposal is a card that you approve or reject, and nothing is written or run until you approve it.

The agent can be wrong. Read Agent cautions and limits before you raise its level.

Where to go next

I want to	Read
Turn it on, read insights, approve a retry, set rules, switch it off	Using the AI Monitoring Agent
Understand levels, approvals, standing authorizations, pauses and quiet hours	Authorization levels and grants
Know what to watch out for	Agent cautions and limits
See what is sent, stored and deletable	Agent data and privacy
Decide whether a retry is safe	When is a retry safe?
Work out why a run failed	Troubleshooting failed task runs

Documentation