Agent Designer

Overview

Agent Designer is the section of the settings console where account administrators manage Xurrent Agents. Each agent has a set of configurable skills, and each skill can be tuned to the level of automation a team wants.

Two agents are available: the Triage Agent and the Closure Agent. The console shows each agent's state, so an agent that is enabled and active is distinguishable from one still awaiting onboarding.

Skills and action modes

Every skill runs in one of two modes, set per skill:

• Suggest posts a recommendation for a person to review. Nothing changes on the record until a human acts.
• Auto Confirm performs the action directly.

Because the mode is set per skill and can be tuned per team, a team can run duplicate detection in Auto Confirm while keeping impact assessment in Suggest.

Confidence thresholds

Each skill also carries a confidence threshold, set with a slider. The threshold is the minimum confidence the agent must reach before it takes the action. Raising the threshold reduces the number of actions and raises their precision; lowering it does the reverse.

Assigning an agent to work

An agent can be added to a team like any other member. Once it is a team member, a human team member can assign it to individual requests, and it appears in the Member picker on requests and workflow tasks alongside people.

Agents are assigned manually by team members. Rule-based or trigger-based assignment is not available.

Dashboard metrics

The Agent Designer dashboard reports Active Agents, Runs This Week, Avg Success Rate, and Time Saved/Week.

Avg Success Rate

The share of Triage Agent runs that took at least one action and were not corrected within the correction window, currently seven days. Runs that took no action are excluded from the calculation. An action a person corrects inside the window counts as a failure.

Time Saved/Week

The sum of an administrator-configurable minutes-saved credit for each uncorrected action across the week. It is displayed as an estimate.

An action corrected within the window contributes zero time saved. Any action type without a defined credit also contributes zero, so adding a new action type cannot silently inflate the figure.