How to Set Up AI Agents That Propose Fixes and Wait for Approval

What's in the demo

• Settings, AI Agents: an alert triage agent for all clients, running in Shadow, with "2 findings to review"; a patching agent showing its next scheduled run; an organization-only help desk pilot marked Inactive because no partner-wide help desk agent exists yet
• New agent: Help desk, owned by all organizations, then the three modes. Off runs nothing. Act executes a fixed set of allowed operations without a click, with no automatic rollback, so you have to tick an acknowledgement first. Shadow observes and proposes, and a human approves everything
• What it does: Use recommended (for a help desk agent, fixes like a service restart or a disk cleanup)
• Safety and oversight: protected services, paths and registry keys, plus limits on runs per hour and daily spend
• The review card, including "Executes unattended", then Create with Start enabled left off
• The triage agent's settings: protected resources, limits, a daily sweep at 6 AM Central (disk pressure, stale agents, pending reboots) and the graduation panel that tracks each kind of fix
• AI Agent Runs: alert, schedule and manual runs with status, trigger, verdict and Show cost
• One run end to end: LSL-WS-03 pegged at 97% CPU by a stuck print spooler, the proposed spooler restart as a request in Approvals ("Review in Approvals"), every tool the agent called, and the linked approval
• Fix memory: the spooler restart's track record (13 of 14, shared across all clients), a macOS fix demoted after failing twice in a row, and reviewed steps your team signed off on
• AI Impact: Refresh rebuilds the numbers from the runs. Over 30 days: alerts judged, alerts flagged as noise, fixes a person approved that were carried out, an estimated time saved (counts times weights you can edit), model spend, and a breakdown by client

Questions this video answers

Does a Breeze AI agent change devices on its own?

Not in shadow mode, which is where every agent should start. In shadow it investigates and proposes. Risky fixes become requests in Approvals, and nothing runs until a person approves one. Approving runs it for real, not as a dry run.

What does act mode do?

Act executes a fixed set of allowed operations without a click, one device per run, up to the run's action cap. Verification can still fail, and there is no automatic rollback, so Breeze asks you to acknowledge that before you can choose it.

How do I keep an agent away from critical systems?

Protect them. Services, paths and registry keys on the protected list are denied before a proposal reaches anyone. Limits cap how often the agent runs and what it spends in a day.

How do I turn AI agents on, and off?

AI agents run only on a server started with BREEZE_AI_AGENTS_ENABLED=true; without it, the AI Agents page says agents are disabled on this server. Switching an agent off stops it, and whoever runs the server can stop every agent at once. New agents are created switched off.

Where do I see what an agent did and what it cost?

AI Agent Runs lists every run with its status, trigger and verdict, and Show cost adds what each one spent. Open a run for its summary, the actions it proposed, every tool it called and the linked approvals. AI Impact rolls it up over 7, 30 or 90 days, by client.