Braze Agents in Agent Console

Braze Agents are AI-powered helpers you can create inside Braze. Agents can generate content, make intelligent decisions, and enrich your data so you can deliver more personalized customer experiences.

Watch this video for an overview of Braze Agents in Agent Console.

Why use Braze Agents?

Braze Agents help your team deliver personalized experiences—without adding extra work. They understand context, make decisions, and take action toward a goal, rather than only responding to a single prompt.

In practice, agents can automatically create message copy—like subject lines or in-product text—so every customer gets communication that feels tailored to them. They can also adapt in real time, routing people through different Canvas paths based on preferences, behaviors, or other data.

Beyond messaging, agents can enrich your catalogs by calculating or generating product and profile field values, keeping your data fresh and dynamic. By taking on repetitive or complex tasks, they free your team to focus on strategy and creativity instead of manual setup. Braze Agents act more like collaborators than background processes—helping you solve problems and deliver impact at scale.

When to use Braze Agents versus other BrazeAI features

Use agents when you need to reason over one user’s context at send time—including free text or a judgment call. For example, if an agent knows a particular user’s favorite ice cream flavor is chocolate and their favorite topping is gummy bears, it can write push copy for that combination as the user passes through the Canvas.

Braze Agents reason over the context you pass in. They don’t learn from whether the output worked, so they don’t improve on their own from outcomes.

Use other BrazeAI features when you need learning from measured results:

If you want to	Use instead
Score how likely a user is to churn or act	Predictive Suite
Pick items from a large catalog	Item recommendations (then use an agent to write copy around those items)
Improve results against a metric over time, or decide offers	Decisioning Studio
Draft or refine content while you build in the dashboard	BrazeAI Operator

For a full comparison of BrazeAI features, see BrazeAI.

Features

Features for Braze Agents include:

• Flexible setup: Use a Braze-provided LLM or connect your own AI model providers (such as OpenAI, Anthropic, Google Gemini, or Databricks Mosaic).
• Seamless integration: Deploy agents directly in Canvas steps or catalog fields.
• Testing, logging, and version history: Preview your agent’s output by testing with sample inputs before you launch. View logs for each time the agent runs, including the input and output for that run. Use the Version history tab to review past versions and inline diffs of instruction changes.
• Usage controls: Daily limits help manage performance and costs.
• Agentic Standards: Define rules and rulesets in Agent Console, then run an agentic evaluation before you launch a campaign or Canvas to check brand guidelines, organizational conventions, and technical requirements.

About Braze Agents

Agents are configured with instructions (system prompts) that define how they behave. When an agent runs, it uses your instructions along with any data you explicitly pass in to generate a response. They cannot access user data beyond what you configure—Liquid variables, Agent context selections, Canvas context variables, and Context step values. Agents do not search profiles or warn when data is missing. See What data agents receive.

Key concepts

Term	Definition
Model	The agent’s “brain,” in this case a large language model (LLM). It interprets inputs, generates responses, and performs reasoning. A stronger model (trained on more relevant data) makes the agent more capable and versatile.
Instructions	The rules or guidelines you give the agent (system prompt). They define how the agent should behave each time it runs. Clear instructions make the agent more reliable and predictable.
Context	Data passed into the agent at runtime, wherever it is deployed, such as user profile fields or catalog rows. This input provides the information the agent uses to generate outputs.
Canvas context variables	Temporary pieces of data you can create and use within a user’s journey through a specific Canvas.
Output variable	The output the agent produces when used in Canvas steps. Output variables store the agent’s result to personalize content or guide workflow paths. Output variables can be a string, a number, or a boolean data type.
Invocation	A single run of the agent. This counts against your daily limits.
Output format	The predefined data structure of the agent’s response.
Knowledge sources	A type of agent context used to retrieve data from a catalog more accurately than if the catalog is referenced directly in the agent’s instructions.
Conversational Agents	A channel for multi-turn user conversations. You define conversational workflows with tools so users can complete tasks over SMS, RCS, WhatsApp, or web chat.

Limitations

The following limitations apply:

• Each agent has a default daily invocation limit of 250,000 runs, which can be increased up to a maximum of 1,000,000 runs per day. Contact your customer success manager if you’re interested in increasing this limit.
• Agent Console shows a Daily action credit cost limit for each agent—the estimated maximum credits per day based on your model’s per-invocation credit ratio and the daily invocation limit. See Daily invocation and credit limits.
• By default, each run must complete within 20 seconds. After 20 seconds, the agent returns a null response where it is used.
  • If your agents consistently time out, contact your Braze account manager to increase this limit.
• Input data is limited to 25 KB per request. Longer inputs are truncated.

Best practices

Target high-value use cases where agents can drive the biggest return on investment (ROI), and choose audiences who are likely to respond. A smaller, high-opportunity audience often outperforms a large audience with low opportunity—for example, retargeting users who searched recently but did not convert, rather than sending agent-generated copy to your entire user base.

To validate ROI before scaling, use an Experiment Paths step to send only part of your audience through an Agent step. When a small-scale test looks good, scale the agent to your full target audience and raise the daily invocation limit so invocations are not capped mid-send. Confirm you are comfortable with the estimated credit consumption before scaling to your full audience. For more deployment guidance, see Deploy custom agents.

Error handling

If the connected model returns a rate limit error from the LLM provider during a Canvas Step Agent or Catalog Agent invocation, Braze continuously retries the request using exponential backoff.

For other failures (such as a timeout or invalid API key), the Canvas Step Agent output is set to null unless the agent has fallback values configured in Agent Console (Canvas Step Agents only). Catalog agents do not retry non-rate-limit failures. If an agent reaches its daily invocation limit, Braze applies configured fallback values when present; otherwise the output is set to null.

Rate limit errors, model unavailability, and daily invocation limit failures do not consume Braze credits. Timeouts do consume credits. See When credits are consumed.

When many users enter an Agent step at once, processing may take longer because of invocation flow controls. Configure fallback values in Agent Console for Canvas Step Agents so users still receive output when an invocation fails, or use default Liquid values in downstream Message steps.

How is my data used and sent to Braze-provided LLMs?

In order to generate AI output through Braze AI features that Braze identifies as leveraging Braze-provided LLMs (“Output”), Braze will send your system prompt or any other input, as applicable (“Input”) to the Braze-provided LLM. Data sent to the applicable Braze-provided LLM is not used to train or improve the Braze-provided LLM. Between you and Braze, Output is your intellectual property. Braze will not assert any claims of copyright ownership on such Output. Braze makes no warranty of any kind with respect to any AI-generated content generally, including Output.

The Braze-provided LLM for Braze Agents, identified as “Auto”, uses Google Gemini models. Google retains Inputs and Outputs submitted through Braze for 55 days, after which the data is deleted.

Next steps

Now that you know about Braze Agents, you’re ready for the next steps:

• Create custom agents
• Deploy custom agents
• Agentic Standards
• Conversational Agents