How Many AI Agents Do You Need?
We're several months into 2025, and the excitement around AI agents hasn’t cooled down.
Nearly every tech vendor is busy launching their own AI agents, or capabilities allowing you to build custom agents. Salesforce made headlines by announcing their platform will host one billion agents by the end of this year. Yet, amidst all the buzz, there’s one question I’ve noticed is rarely discussed: How many AI agents does your organization really need? Is it one? Ten? A hundred? A thousand?
The Case for a Single Agent
In November 2022, OpenAI rolled out ChatGPT—a chat-based interface built on its latest large language model (LLM) at the time: GPT-3.5. This kicked off a race to offer the single best-performing LLM across a wide variety of tasks.
Starting with just one model is practical, similar to starting with just one agent. When selecting this first agent, your goal likely isn't perfection in one particular area, but rather good performance across many tasks. Think of it like hiring the first employee at a startup—you'd choose someone versatile who can wear multiple hats, rather than a specialist who excels only in one narrow area.
The Case for Ten Agents
We quickly saw LLM providers move from offering a single model to several models. Through its ChatGPT interface, OpenAI now provides access to GPT-4.5, o4, and o4-mini, all with slightly different strengths.
The conversation around AI agents is trending in this direction. Rather than having one general-purpose agent, vendors now offer specialized agents designed for specific tasks. It’s like a startup reaching a stage where it sets up distinct departments like Finance, HR, and Sales, each with dedicated staff.
Many vendors with agentic solutions are already here, offering or announcing multiple specialized agents. Adobe provides several CX agents, including an account qualification agent, an audience agent, and a content production agent. Similarly, GrowthLoop highlights eight different agents on its website.
The Case for One Hundred Agents
Reaching one hundred agents within an organization might be closer than it seems. Imagine deploying 10 agents for Marketing, 10 for HR, and another 10 for Finance—you'll quickly hit triple digits.
As a company grows and matures, adding layers of management typically becomes necessary. Companies differ in their philosophies about how many direct reports a manager should handle, and despite recent trends toward flattening organizations (often called "the great flattening"), it's unlikely we'll ever see a scenario where everyone reports directly to a single CEO.
Something similar is emerging in the world of AI agents. To handle multi-agentic systems, we're seeing the rise of "supervisors"—agents responsible for managing and coordinating other agents. These supervisors require distinct skills, mainly focused on assigning tasks to “action agents” efficiently based on specific criteria, like performance or cost.
For example, a marketing supervisor agent aiming to launch a new campaign might delegate tasks to its specialized agents: one to analyze potential customer segments, another to craft compelling messages (both creative and text), and a third to determine the optimal communication channels.
The Case for One Thousand Agents
Enterprises often employ thousands of people, many with highly specialized knowledge and skill sets. Could we see a similar situation with AI agents? In a future where specialized agents outperform general-purpose ones, this scenario might indeed happen. It could even be necessary to reach the billions of deployed agents that Salesforce predicts.
However, personally, I doubt we’ll reach that scale soon. AI agent systems powered by LLMs are rapidly evolving and remain immature. Right now, it's challenging enough for organizations to effectively manage even a fleet of 100 agents, let alone 1,000.
LLM Evolution: A Model for the Future of Agents?
One model to rule them all? Steve Jobs was deeply committed to simplicity, both as Apple's guiding principle and in his personal approach to leadership and design. He believed genuine simplicity—stripping products and processes down to their essentials—was the key to creating intuitive, user-friendly technology and effective organizations.
Sam Altman, OpenAI’s CEO, seems to share a similar philosophy. Earlier this year, he tweeted that OpenAI aims to "make AI just work," signaling a future where users won't need to choose between multiple models with confusing, abstract names.
Netflix recently provided another compelling example in a blog about their recommendation system—one of the key reasons the company is valued above $500 billion today. Netflix transitioned from numerous specialized recommendation models to a single, more comprehensive foundation model for personalized recommendations.
All of that to say, we may not just see an ever expanding number of agents in organizations. Contraction could very well be seen at some point in the future!
Charting Your Path to AI Agent Maturity
Organizations today range from just a handful to thousands of employees. In the coming months, even the largest enterprises are unlikely to deploy more than a few dozen AI agents in production. The path toward maturity might look something like this:
Start simple: Deploy a single agent for one specific use case.
Expand usage: Gradually apply this single agent across several related uses.
Introduce specialization: Deploy multiple agents, each tailored for specific roles or tasks.
Orchestration: Add the first supervisor agent to manage and coordinate multiple specialized agents.
Optimize and evolve: Continuously refine the overall agentic system, leading to either expansion with additional agents or consolidation around fewer, more capable ones.
🔑 The future isn’t about deploying the most AI agents—it’s about deploying the right agents effectively.