name: decisioning-studio-design-agents source_url: >- https://braze-inc.github.io/braze-docs/_user_guide/brazeai/decisioning_studio/design_agents indexed_at: '2026-04-05' keywords:
- decisioning
- agent
- metric
- dimensions
- optimization
- experiments
- constraints
- actions
- audience
- conversions triggers:
- design a decisioning agent
- configure success metrics
- set up action dimensions
- define agent constraints
- optimize customer engagement
★ Insight ─────────────────────────────────────
Topic files in this codebase are "atomic knowledge units" stored in skills/{id}/references/*.md. They're designed for fast lookup at the Default depth (Sonnet), so stripping Jekyll/Liquid templating and preserving dense, scannable structure is exactly right — the MCP semantic search needs clean prose, not template noise.
─────────────────────────────────────────────────
Designing Decisioning Agents
A decisioning agent is a custom configuration in BrazeAI Decisioning Studio that optimizes a specific business goal by experimenting with and learning which combinations of actions work best for each customer.
Core Concepts
| Term | Definition |
|---|---|
| Decisioning agent | Custom configuration targeting a specific business goal, defined by its success metric, dimensions, and options. |
| Success metric | The business metric to optimize (e.g. revenue, conversions, ARPU, CLV). The agent maximizes this through its actions. |
| Dimensions | The types of levers the agent can pull — e.g. offer, subject line, creative, channel, send time. |
| Action bank | The specific options available for each dimension lever. Defines the full universe of possible agent actions. |
| Constraints | Rules that limit agent actions to respect business requirements (e.g. geo-eligibility rules, budget caps). |
The agent can only take actions explicitly configured in the action bank. All possible behaviors are combinations of what you put there.
Four Design Elements
1. Success metric ("the goal") What outcome should the agent maximize? Use real business results — revenue, conversions, ARPU, customer lifetime value — not proxy metrics like clicks or opens.
2. Audience ("the who") Who will the agent engage? Options include all customers, a segment (e.g. loyalty members), or a lifecycle cohort (e.g. recent purchasers, at-risk subscribers).
3. Action bank ("the what") Define the dimensions and the specific options within each. The agent experiments across combinations to find what works best per customer.
4. Constraints ("the how") Define rules the agent must follow — geography restrictions, budget limits, frequency caps, or eligibility rules.
Best Practices
- Choose a success metric that directly aligns with business objectives, not vanity metrics.
- Prioritize dimensions most likely to move the needle on your success metric.
- Select dimension options (e.g. email vs. SMS, daily vs. weekly) based on expected impact.
- Decisioning Studio runs daily experiments automatically — no manual A/B test management needed.
Agent Examples
| Agent Type | Goal | Key Dimensions Tested |
|---|---|---|
| Repeat purchase | Increase follow-up conversions post-sale | Product offers, message timing, frequency |
| Cross-sell / upsell | Maximize ARPU from subscriptions | Messages, send times, discounts, plan offers |
| Renewal & retention | Secure contract renewals, maximize NPV | Renewal offers, discount levels |
| Winback | Reactivate lapsed subscribers | Creative, message, channel, cadence |
| Referral | Drive new account openings via referrals | Emails, creatives, send times, card offers |
| Lead nurturing | Drive incremental revenue, optimize cost per customer | Customer segments, bidding methodology, bid levels, creative |
| Loyalty & engagement | Maximize purchases by new loyalty enrollees | Email cadence, offers, messaging |
BrazeAI learns the best combination for each individual customer over time, then orchestrates personalized sends through Braze to maximize the configured success metric.
★ Insight ─────────────────────────────────────
The tabular agent examples consolidate the {% tabs %} block (a Jekyll UI component) into a single scannable table — this is idiomatic for topic files since the MCP search indexes plain text. The four design elements map directly to the "goal/who/what/how" framework in the source, preserving the mental model without the rhetorical questions.
─────────────────────────────────────────────────