Loomi, Composer, and Athena: The Race for the AI Marketing Agent Has Started. It Will Be Won by What the Agent Can Call.

AI marketing agents like Bloomreach’s Loomi, Klaviyo’s Composer, and Zeta’s Athena are no longer science fiction. They’re shipping now, and they promise to automate campaign creation from a single prompt. The race is on, but the real differentiator isn’t which agent writes the best copy. It’s which one can call the best deterministic tools to render brand-locked content and measure real revenue impact. That’s where most agentic AI projects will fail, and where the few will win.

The Coding Harness War Already Happened

Before marketing had AI agents, developers did. Tools like Claude Code, OpenAI’s Codex, and Cursor promised to write code from natural language prompts. But the winning architectures weren’t the ones that generated the most creative code. They were the ones that integrated tightly with deterministic systems: compilers, test runners, version control, and typed APIs. The agent didn’t ship the final code, it orchestrated calls to tools that did, with predictable, auditable outputs.

That same pattern is repeating in marketing. Just as developers needed tools that produced reliable, testable code, marketers need tools that generate on-brand, measurable content. Freeform AI generation fails here because it’s unpredictable, hard to govern, and nearly impossible to attribute. The future isn’t AI writing the final email. It’s AI assembling the campaign, then calling a content service to render the final, revenue-optimized experience deterministically.

What Marketing Agents Do Today

Bloomreach launched Loomi Marketing Agent into general availability on June 2, 2026, positioning it as a system that enables “autonomous campaign creation and optimization” using real customer data to personalize each campaign (Business Wire, June 2, 2026). It’s designed to build full lifecycle campaigns from a single prompt, workflow, channels, content, and performance monitoring.

Klaviyo’s Composer acts as a chat-style agent inside the platform, generating subject lines, preview text, and full email content from a description. Early practitioner feedback, like a July 6, 2026 review from a real email marketer, highlights concerns about output quality and brand control, exactly the gap a deterministic content layer fills.

Zeta Global introduced Athena as a “superintelligent natural language AI agent” that personalizes the marketer’s workspace and operates the platform on their behalf (MarTech, October 10, 2025). On January 5, 2026, they announced a partnership with OpenAI and moved Athena’s first agentic applications, Insights and Advisor, into beta, showing how marketing clouds are owning the harness while sourcing model power externally.

These aren’t just assistants. They’re full campaign-building agents. And they’re here.

Gartner Predicts the Agency Will Be Cancelled

Here’s the tension: Gartner predicts that by the end of 2026, 40% of enterprise applications will be integrated with task-specific AI agents, up from under 5% in 2025 (Gartner, August 26, 2025). That’s explosive adoption.

But in the same breath, Gartner also predicts that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs and unclear business value (Gartner, June 25, 2025).

Why such a split verdict? Because most AI agents today are built on freeform generation, creative, yes, but unmeasurable and non-deterministic. If you can’t verify the output, can’t attribute revenue to it, and can’t ensure brand consistency, it doesn’t matter how fast the agent works. Finance will kill it.

The winners will be the agents that call deterministic services, tools that render pixel-perfect content, lock brand fidelity, and report back revenue per block, per variant, per channel.

The Deterministic Content Layer

So what should a marketing agent call? Not another AI model. It needs a service with three qualities:

• Deterministic rendering: The final image is built from brand-approved templates and assets, not generated on the fly. No hallucinations, no off-brand colors, no broken layouts.
• Single tag integration: One dynamic tag that works across email, SMS, RCS, and app notifications. The agent drops in the URL, the service handles the rest, personalized at open time, per recipient.
• Revenue feedback signal: Block-level RPM and CTC attribution that flows back to the agent, so it can optimize not just for opens, but for revenue.

This is the layer we’re building at Zembula. We’re not competing to be the AI agent. We’re the experience layer any agent can call. When a marketer prompts their AI to build an email, the agent can call Zembula’s MCP, drop in a single content block tag, and know that the final render will be brand-perfect, personalized at open, and tied to a revenue readout.

Our composition engine uses no generative AI in the final image. It’s deterministic, fast, and auditable. And because we support open-time decisioning, the content adapts to what the recipient actually sees, not what we assumed they’d want when we hit send.

Plus, our Block-level RPM and CTC attribution gives agents the feedback loop they need to optimize. Most platforms can’t tell you which part of the email drove revenue. We can, down to the individual email block.

This isn’t hypothetical. We’ve seen Smart Banners drive returns as high as 41x (J.Crew) and 30x (Pair Eyewear, at 99.939% significance). On average, brands see 15-20% incremental revenue from just two Smart Banner slots, validated through longitudinal holdout testing. For the full data, see our 2025 email performance benchmark report.

AI Proposes, Humans Approve, Tests Verify

The most durable AI governance model isn’t full autonomy. It’s AI proposes, humans approve, tests verify. That’s the framework we laid out in a recent post, and it’s already proving resilient.

Agents will propose campaign structures, copy variants, and audience segments. But the final creative, especially anything customer-facing, should be rendered deterministically from approved templates. And every change should be tested against a holdout group to prove real channel impact, not just vanity metrics.

This is how you avoid the 40% cancellation rate. By making AI outputs measurable, brand-safe, and revenue-optimized.

Our D.A.V.E. (Dynamic Automated Variant Engine) already operates on this principle, and it’s a preview of the worker agent economy Gartner predicts, where 15% of day-to-day decisions will be made autonomously by 2028 (Gartner, January 15, 2026).

Key takeaways

• The race for AI marketing agents is already underway, with Loomi, Composer, and Athena leading the charge.
• But the winning agent won’t be the one with the best generative AI, it’ll be the one that calls the best deterministic tools.
• Agents that generate final creative directly will fail due to brand risk and lack of measurement.
• Agents that call a deterministic content service, with brand-locked templates, single tag integration, and revenue feedback, will thrive and scale.
• Zembula is not building a competing agent. We’re building the experience layer any agent can call, ensuring content is on-brand, personalized, and measurable.
• The future of AI in marketing isn’t autonomy, it’s orchestration.

For more on how deterministic AI generation works in practice, read Image Personalization Email at Scale: Why AI Generation Breaks (and the Math to Prove It). To understand the full stack, see What is a Composition Engine? And to see how email stacks up as a performance channel, read Email Is a Performance Marketing Channel, and the Math Proves It.