Agentic commerce needs purpose-built infrastructure, not workarounds
The rise of AI agents that can autonomously research, negotiate and purchase on behalf of users promises to revolutionize commerce.
But as companies rush to build agentic shopping experiences, most are relying on makeshift solutions: browser automation that scrapes websites, point-to-point API integrations or legacy eCommerce platforms never designed for autonomous agents.
While these approaches work for demos, they create a fragile patchwork that can't handle the scale, complexity and trust requirements of true agentic commerce.
Here's why the infrastructure matters more than the interface, and what it takes to build commerce systems that agents can actually rely on.
Agentic commerce is here — but the infrastructure isn’t
AI agents are transforming how people discover, decide and buy. In this new paradigm, buyers don't just browse or click — they delegate. Agents act on their behalf, handling everything from product research to price negotiation to final purchase.
From OpenAI's GPT-based agents to specialized agentic startups like Operator and Comet, the movement is gaining momentum. Gartner research suggests that AI agents could account for 20% or more of eCommerce traffic within five years. But most implementations today rely on either:
Browser agents (scraping HTML, automating clicks)
Commerce stacks (Shopify, Salesforce, Magento)
Custom tool integrations, built one seller or endpoint at a time
They all share one critical flaw: they weren't designed to scale autonomous commerce operations. What works for a demo with a few merchants breaks down when agents need to operate across thousands of sellers, handle complex transactions, and maintain trust at scale.
A patchwork of web pages and legacy stacks
Commerce today is fragmented by nature.
Thousands of isolated platforms, each with its own rules, formats and interfaces. Amazon owns ~40% of U.S. eCommerce, but operates as a closed ecosystem with proprietary APIs and seller requirements. Walmart, Target, and other major retailers split much of the remaining market, each running on distinct technology stacks and seller infrastructures that don't communicate with each other.
Meanwhile, thousands of brands operate independently on Shopify, Magento, or Salesforce Commerce Cloud, creating a long tail of siloed commerce experiences. Each platform has evolved its own data schemas, checkout flows and business logic — optimized for human shoppers, not autonomous agents.
For an AI agent trying to operate across this landscape, the challenge is real:
No consistent data model across platforms
No unified product taxonomy or attribute standards
No shared checkout, payment or compliance logic
Varying API capabilities and rate limits
What you get is a web of bespoke formats, inconsistent business logic and fragile connections. It's hard enough for humans to navigate. It's nearly impossible for agents to operate reliably at scale.
Why browser agents break at scale
Browser-based agents (like Operator, Comet, Manus, etc.) offer impressive demos. They "browse like humans" using headless browsers, simulating clicks and reading HTML to navigate websites and complete purchases.
But they don't connect to underlying commerce systems — they scrape. That means:
Loading...
Browser agents treat symptoms, not causes. They're essentially building a sophisticated workaround for broken infrastructure rather than fixing the infrastructure itself. When a website changes its layout, the agent breaks. When a merchant updates their checkout flow, integration fails. When traffic spikes, scraping becomes unreliable.
In concrete terms, this leads to:
Brittle integration: Every website redesign potentially breaks the agent.
Poor performance: Rendering full web pages is slow and resource-intensive.
Limited data access: Can only see what's visible to human users, missing rich product data.
No transaction guarantees: Cannot ensure order completion or handle edge cases.
Compliance gaps: Cannot access or enforce merchant policies, SLAs or regulatory requirements.
They're a clever layer on top of broken infrastructure. And broken infrastructure doesn't scale.
Point-to-point tools aren’t the answer either
Some agentic implementations take a different approach, using Model Context Protocol (MCP) or plugin-based tool calls to directly access commerce APIs.
This seems more sophisticated than browser scraping, but comes with its own scaling problems, and becomes a game of API whack-a-mole:
Custom connectors: Every merchant requires a unique integration, each with different authentication methods, data formats and business logic.
Maintenance: Any taxonomy change, API update or policy modification requires rebuilding tools across the entire network.
High dependency: These agents depend on external systems that weren't designed to be "agent-aware," leading to unexpected failures.
Limited functionality: Most merchant APIs are designed for basic operations, lacking the rich functionality agents need for autonomous decision-making.
For example, an agent trying to purchase office supplies might need separate integrations for Staples, Office Depot, Amazon Business and dozens of specialty suppliers.
Each integration requires different authentication, uses different product categorization and offers different capabilities. When suppliers change their APIs (which happens frequently), the entire network becomes unreliable.
In short: point-to-point integration = point of failure at scale.
Mirakl Nexus: a platform built for agents
Even popular commerce platforms like Shopify, Salesforce Commerce Cloud and Amazon have critical limitations for agentic workflows. These platforms were designed for human shoppers browsing individual storefronts, not for AI agents operating across multiple sellers and complex supply chains.
Loading...
The platform gap is real: Traditional eCommerce platforms work for brands, not agent-driven ecosystems.
Mirakl Nexus is an agentic commerce operating system that provides:
Real-time APIs for instant access to product data, pricing and inventory across thousands of sellers.
Native orchestration for checkout, payments, refunds, tax calculation and SLA enforcement.
Trust and compliance enforcement at the platform level.
Modular architecture compatible with any agentic interface.
Unified data model that standardizes information across the entire ecosystem.
Mirakl Catalog Transformer, a part of Mirakl Nexus, solves the data quality problem. Most sellers onboard with messy, incomplete product content. This LLM-native engine transforms inconsistent seller data into structured, agent-ready content by:
Parsing descriptions and images to understand products
Filling missing attributes (size, color, material, brand)
Normalizing categories and converting units
Optimizing for agent discoverability and compliance
This accelerates seller onboarding from weeks to days while ensuring agents get reliable, standardized data they can act upon.
Mirakl Nexus abstracts commerce fragmentation and offers a standardized interface built for the scale, trust and speed that agents require.
Mirakl Nexus in action: from raw input to agent-ready listings
Most vendors provide minimal product information. Mirakl Catalog Transformer uses advanced LLMs to transform sparse data into rich, structured content that agents can reliably interpret and act upon.
Loading...
Now the product is searchable by LLM agents, compatible across multiple marketplaces and immediately ready for activation.
Beyond text analysis, Catalog Transformer uses computer vision to extract additional product details from images — identifying materials, colors, style elements and usage contexts that sellers often forget to mention. This multimodal approach ensures agents have comprehensive product understanding, not just basic descriptions.
The result: products become immediately searchable by AI agents, compatible across multiple marketplaces, and ready for autonomous purchasing decisions with rich context that goes far beyond what human shoppers typically see.
Bottom line: Agentic commerce needs a platform, not a patchwork
Browser agents are clever. Point-to-point tools are flexible. Legacy stacks are mature. But none of them offer the scale, trust and structured intelligence needed for real agentic commerce.
To move from demo to deployment, we need:
✅ A commerce OS (Mirakl Nexus)
✅ Rich, structured, multimodal product data (Catalog Transformer)
✅ Open, modular, agent-ready infrastructure
This is how we go from fragmented commerce to composable agentic ecosystems.
Because the future doesn’t need more workarounds.
It needs platforms built to scale.
