How Many AI Agents Do You Need? Walmart Says… Four
A few weeks ago, I shared thoughts on the number of agents we might see in organizations — from just one to potentially thousands. After all, Salesforce has predicted a billion agents deployed on Agentforce in 2025.
Back then, the question was: How many agents does your organization actually need?
Walmart’s latest announcement makes me think the more relevant question might be: How many agents should a user ever see?
Walmart’s Four Super Agents
Walmart introduced a strategy built around four “super agents”, for customers, associates, partners, and developers.
Each acts as a single entry point. Behind the scenes, multiple specialized agents handle everything from supplier onboarding to sales trend analysis to benefits navigation.
The reason? Too many visible agents can overwhelm users. By orchestrating the work in the background, Walmart keeps the experience simple: one interface, without showing the complexity underneath.
The Example From GPT-5
OpenAI’s GPT-5 release last week points in the same direction.
In the past, you picked between o4, o4-mini, or gpt-4.5. As Sam Altman suggested earlier this year, OpenAI aims to lower the barrier for users by deciding which model will deliver the best result—rather than putting that choice on the user.
The result: the user sees one “face” of the system, while model selection is still happening, but in the background.
How Many Do You See vs. How Many You Have
Software has operated this way forever. Nobody knows (or cares?) how many APIs power their favorite app. Architects decide the right granularity and wrap it in a clean, unified interface.
AI architects will follow the same path:
Determine how specialized each agent should be
Orchestrate them behind the scenes
Keep the number of visible entry points small
As a result, users interact with a handful of supervisor agents—or “super agents,” as Walmart calls them—while dozens, or even hundreds at some point, of specialized agents run quietly in the background.
🔑 The key takeaway from the previous post still holds: The future isn’t about deploying the most AI agents — it’s about deploying the right agents effectively.