Managing Memory for AI Agents

Chapter 3. Some Economics of Agents, Model Usage, and Selection

LLMs serve as the cognitive engines that power AI agents—just like how different car engines are built for different purposes, from ATVs to school buses. Extending this framework, outlined in What Are AI Agents? When and How to Use LLM Agents (O’Reilly), selecting the right LLM for your agent isn’t just about performance—it’s also about a trade-off between capabilities and economics. The foundation model you choose will directly affect not only what your agent can do but also how much it costs to run, how efficiently it operates, and ultimately, whether your project makes financial sense. Understanding these economic trade-offs is crucial for anyone building AI agents, whether you’re a startup watching every dollar or an enterprise looking to scale.

The Economics of Agent Adoption

The first thing to note about AI agents is that as they expand in their capacity to access and work with more tools—through protocols like Model Context Protocol (MCP) or through functional calls within LLMs themselves—the end user benefits more from the agent. But as with everything in economics, there is a trade-off. As the complexity of the task an agent is provided increases, so too does the marginal cost of a given action—whether we measure that cost by the hours it takes to complete a task, the cost of API calls, or the ability of the underlying agent to solve the niche tasks.

The chart shown in Figure 3-1 plots task complexity ( ...