Agents

Agents are the foundation of AI Decisioning. Each agent defines who to target, what to optimize for, and when messages can be sent.

Learning objectives

After reading this article, you’ll know how to:

• Create a new agent
• Configure your agent
• Understand how agents use messages
• Manage offers to factor cost and value into decisions
• Monitor agent health and pipeline runs
• Set up alerting for agent issues
• Translate content variants across locales
• Monitor and manage active agents

Overview

An Agent in AI Decisioning combines your audience, messages, and goals into a single adaptive campaign.

Agents use reinforcement learning to decide which message, timing, and channel will drive the best outcome for each user. They operate continuously, learning from real user outcomes such as clicks, purchases, or sign-ups.

Each new message helps the model understand which creative and timing combinations perform best for different audiences.

Create a new agent

1. Go to AI Decisioning → Agents.
2. Click + Add agent.

   

3. Enter a name that reflects your campaign’s purpose, such as Win-back customers or Subscription renewals.
4. Choose a target audience from Customer Studio → Audiences.
5. (Optional) Add a holdout group to measure incremental lift (for example, 20%).
6. Click Add agent.

Configure your agent

After creation, click your agent to open the Configuration tab.

From here, you can set audience eligibility, messages, goals, Smart Suppression, and scheduling preferences.

Step 1: Set audience eligibility

This section shows the audience your agent targets and any experimental setup, such as a holdout group.

Step 2: Add and manage goals

Goals define what your agent optimizes toward.

Each goal represents a measurable event (like a purchase or click) and includes a priority level—helping AID decide which outcomes matter most.

To add a goal:

1. In your agent’s Configuration tab, under Goals, click + Add goal.

   

2. Configure the goal.

   • Select an event model (for example, Purchase event, Viewed Product, or Redeemed offer).
   • (Optional) Add a filter to refine the event, such as Performed → Where property is campaign = Summer Sale.
   • (Optional) Select related collections (for example, Live Offers, In-stock Products) to attribute conversions to catalog items.

   

3. Click Continue to move to the Impact step.

   • Choose whether to include this goal in decisioning or track it for measurement only.
   • Select a classification:
     • Best — Primary success metric
     • Very good / Good — Positive supporting outcomes
     • Bad / Very bad / Worst — Negative outcomes such as Unsubscribe
   • (Optional) Add a weight column to scale results by value (for example, Price or Order Value).

   

4. Click Continue to finalize.

   • Enter a clear, descriptive goal name (for example, Redeemed offer).
   • Click Add goal to save.

   

Step 3 (optional): Configure scheduling

Use the Scheduling panel to control how often and when your agent sends messages.

Step 4 (optional): Enable Smart Suppression

Smart Suppression limits low-impact sends using your agent’s learning data.

It predicts which messages are most likely to drive incremental conversions and suppresses the rest.

1. In Configuration, scroll to Smart suppression (optional).
2. Toggle On.
3. Adjust the Only send top [%] slider.
4. Choose a goal metric to optimize for.
5. Review the performance curve once enough data is collected (typically after 2 weeks).

Step 5 (optional): Advanced settings

Use Advanced settings to connect your agent with your organization’s downstream campaign tracking.

Monitor and manage agents

Once your agent is running, open the Overview tab to review live performance and experiment results.

The Overview includes several key panels:

Use this view to monitor how your messages are distributed, which goals are performing best, and how holdout results compare to treatment performance.

Messages

Messages define what an agent can send and where those messages are delivered.

Each message connects an agent to a destination (such as Braze, Iterable, or Salesforce Marketing Cloud). Agents evaluate messages continuously to determine which content, timing, and channel perform best for each user.

Message setup includes:

• Connecting the agent to a destination
• Adding optimized content and variants
• Applying message- and variant-level guardrails

For step-by-step instructions, see Messages →.

Offers

The Offers tab lets you create and manage offers that agents can factor into their decisioning. When an offer is attached to a message, the agent weighs the message's expected value (value minus cost, weighted by conversion likelihood) alongside engagement signals when choosing what to send.

To create an offer:

1. Open an agent and go to the Offers tab.
2. Click + Add offer.
3. Configure the offer:
   • Name: A short, human-readable name (for example, $10 off first order).
   • Description (optional): Additional context about the offer.
   • External ID (optional): A stable identifier in your source system. Required for CSV upserts to match existing offers.
   • Tags (optional): Key-value attributes to help organize and categorize offers.
   • Offer type: Static offers apply the same cost to every user. Column offers read per-user values from warehouse columns.
   • Cost value: A fixed cost associated with each conversion (for example, $10 off).
   • Cost multiplier: A discount rate applied to each conversion's value (for example, 0.2 = 20% discount).
4. Click Add offer.

Exactly one of cost value or cost multiplier must be non-zero.

Once created, attach offers to individual messages from the message Configuration tab. Each message can have at most one attached offer, and not all messages require one.

Health

The Health tab gives you real-time visibility into your agent's pipeline runs, channel delivery rates, and outcome data freshness. From a single view, you can monitor message generation status, diagnose pipeline failures, and catch data issues before they affect optimization.

For full details, see Agent health →.

Localization

The Localization tab lets you translate content variants to a user's preferred language. When enabled, the agent serves localized versions of your message content based on each user's locale.

To configure localization:

1. Open an agent and go to the Localization tab.
2. Toggle on Enable localization.
3. Configure the following settings:
   • User locale field: Select the column from your parent model that contains each user's locale (for example, locale or language).
   • Locales: Define the locales supported by this agent (for example, en-US, fr-FR).
   • Default locale: The locale displayed in the UI and used as the base for all translations. This locale must always be present.
   • Missing locale: Choose what the system should do when a user doesn't qualify for any locale:
     • Use default locale — Send the default locale content
     • Send nothing — Skip the user
     • Raise an error — Flag the issue for review

Alerting

The Alerting tab lets you set up automated notifications for engine run failures, long-running or stalled agents, and unexpected drops in recommendation volume. Every agent comes pre-configured with a default run failure monitor — no setup is required for monitoring to begin. You can add notification channels (Slack, Email, PagerDuty, SMS, or Webhook) and configure severity levels and thresholds.

For full details, see Agent alerting →.

Troubleshooting

What’s next

• Messages: Create and manage message variants →
• Agent health: Monitor pipeline runs and delivery rates →
• Agent alerting: Get notified about run failures and volume drops →
• QA guide: Validate messages before launch →
• Smart Suppression: Control send volume based on predicted impact →
• Tags: Add and analyze message tags →
• Insights: Measure agent and message performance →
• Collections: Attach item recommendations to messages →
• Workspace configuration: Global defaults, channels, scheduling, and feature matrices →