AI Agents FAQs
General
Section titled “General”What are AI Agents?
Section titled “What are AI Agents?”AI Agents are autonomous assistants within SuiteOp that listen to events from your properties (device alerts, reservation changes, task updates) and take actions on your behalf. You configure what each agent can do, which events it monitors, and whether it needs human approval before acting.
Do agents run all the time?
Section titled “Do agents run all the time?”No. Agents are event-driven — they only activate when events are routed to them. Between events, they consume no resources. Each activation creates a session that processes a batch of events.
Can multiple agents handle the same event type?
Section titled “Can multiple agents handle the same event type?”No. Each event type can only be assigned to one agent per organization. This is enforced by a unique constraint to prevent conflicting actions from multiple agents. If you want to change which agent handles an event type, unassign the current agent first, then assign the new one.
What AI model do agents use?
Section titled “What AI model do agents use?”Agents use the same AI infrastructure as SuiteOp’s built-in chat assistant. The system is designed to maximize cache efficiency — the shared domain knowledge portion of the prompt is identical across all agents, so it benefits from prefix caching.
Permissions & Security
Section titled “Permissions & Security”What can agents access?
Section titled “What can agents access?”Agents can only access data and perform actions that their tool permissions allow. When you enable tools for an agent, the system automatically grants the corresponding organization-level permissions. If you disable a tool, the agent loses that permission.
Can an agent do something I can’t?
Section titled “Can an agent do something I can’t?”No. Agents operate within the same permission system as human users. They cannot access data or perform actions beyond what their configured permissions allow.
How do I limit what an agent can change?
Section titled “How do I limit what an agent can change?”Use the three-tier tool permission system:
- Set sensitive tools to Ask so you can review proposed actions before they execute
- Set read-only tools to Auto for efficiency
- Set tools you do not want the agent to use to Disabled
Are agent actions auditable?
Section titled “Are agent actions auditable?”Yes. Every tool call is logged in the action log with the tool name, arguments, result, who approved it (if applicable), and when it executed. Sessions also record the agent’s reasoning and the events that triggered it.
Budgets & Costs
Section titled “Budgets & Costs”How do token limits work?
Section titled “How do token limits work?”Each agent has a daily and monthly token limit. When either limit is exceeded, the agent’s status automatically changes to Budget Exceeded and it stops processing events. Limits reset at the start of each calendar day or month.
What happens when an agent exceeds its budget?
Section titled “What happens when an agent exceeds its budget?”The agent is automatically paused with a Budget Exceeded status. Incoming events for that agent are not processed until you either increase the budget or the limit resets. You can manually reactivate the agent by updating its status to Active after adjusting the budget.
How can I monitor costs?
Section titled “How can I monitor costs?”Each agent has a usage summary showing daily and monthly token consumption, session counts, event counts, and action counts. Review this regularly to right-size your budget limits.
Sandbox & Testing
Section titled “Sandbox & Testing”Does the sandbox use real data?
Section titled “Does the sandbox use real data?”Yes. Sandbox events are processed with the agent’s real tool permissions. If a mutation tool is set to Auto, the agent will execute it for real. To safely test mutations, set them to Ask mode during sandbox testing so you can review each proposed action.
Can I test a paused agent?
Section titled “Can I test a paused agent?”Yes. Sandbox events are processed even when an agent is Paused. This is by design — it lets you create an agent, test it thoroughly in the sandbox, and only activate it for live events once you are satisfied with its behavior.
How is sandbox different from live processing?
Section titled “How is sandbox different from live processing?”Two key differences:
- Sandbox events are processed immediately with no batch delay
- Sandbox sessions are tracked separately and can be filtered independently from live sessions
Memory
Section titled “Memory”Do all agents share memories?
Section titled “Do all agents share memories?”Agents share memories at the organization, property group, property, and reservation scope levels. Any agent with access to the relevant entity can read those memories. The agent scope is private — only that specific agent can access its own agent-scoped memories.
Can I manually add memories?
Section titled “Can I manually add memories?”Yes. You can create memories at any scope level through the Memories section. Human-set memories default to a confidence score of 1.0 and are labeled as “human_set” in the source field.
What happens when a memory is updated?
Section titled “What happens when a memory is updated?”Updating a memory creates a new version. The previous version is marked as “superseded” but preserved in the database. This versioning system ensures you always have a complete audit trail of what information was available to agents at any point.
How many memories can an agent access?
Section titled “How many memories can an agent access?”During each session, up to 50 relevant memories are loaded into the agent’s context. Memories are prioritized by scope (organization first, then property group, property, reservation, and agent) and sorted by most recent within each scope.
Troubleshooting
Section titled “Troubleshooting”My agent is not processing events
Section titled “My agent is not processing events”Check the following:
- Agent status — make sure the agent is Active, not Paused or Budget Exceeded
- Stream assignments — verify the agent is assigned to the event types you expect
- Budget — check the usage summary to see if the daily or monthly limit has been hit
An action failed after approval
Section titled “An action failed after approval”When you approve an action, the system executes the tool call with the original arguments. If execution fails (for example, the target entity was deleted between proposal and approval), the error is recorded in the action log. The action is still marked as “approved” since the human approved it — the failure is in execution, not approval.
Events are being skipped
Section titled “Events are being skipped”Events that fail processing 3 times are automatically skipped to prevent infinite retry loops. Check the session history for failed sessions to understand what went wrong. Common causes include network timeouts, permission changes, or deleted entities.