It's the Tools

Agents are LLMs with reasoning, tool use, and the motivation to iterate long enough to accomplish complex tasks. As foundational models become smarter and capable of sustained thinking, agents will keep improving.

Tools result from a clever insight that an outer loop can inject relevant information into the context at the LLM’s request. Tool use itself doesn’t represent a breakthrough in model ability but highlights our capability to leverage LLMs in novel ways. Conversely, long context, multi-modality, and reasoning are core innovations within the language models.

If agents are poised to replace a meaningful amount of economic activity, there’s a compelling leveraged beta bet to participate in this transition. Any project intending to benefit from this economic transition needs to build relevant assets. However, improving foundational models is highly constrained by CAPEX and specialized talent, and future winners in foundational models are likely already established.

Hence, a startup should focus on the other side of the equation: the tools.

This explains the growing interest in MCP—a protocol designed to turn APIs into tools accessible by any agent. We see hobbyists on platforms like Reddit building MCPs for everything, and established companies releasing MCPs for their products. There’s also evidence of second-order effects, such as companies blocking access to agents and content providers blocking crawlers. The market has responded swiftly, creating MCP marketplaces, platforms, and infrastructure.

Yet, MCP itself isn’t the limiting factor.

If everything I do on my computer has an MCP, instructing my assistant of choice to select the appropriate one for each request will still fail miserably. Trivial tools such as “book a meeting” or “get weather” are compelling for specific scenarios but fall short of replacing a qualified human employee. Effective agents capable of performing complex, prolonged tasks require significant innovation and iterative development in the tools themselves.

To visualize this clearly, consider the physical world. Imagine asking an LLM to describe the steps to take out the trash. Most wouldn’t be surprised if a model could view an image and provide detailed instructions. However, for AI to accomplish this goal, it requires access to a robot—the bridge between instruction and action.

Yet, this gap isn’t limited to the physical realm. Numerous essential tools remain absent in the digital world. For example, costly desktop CAD software used by engineers in architecture, mechanical design, and circuit design is typically proprietary, lacks APIs, and wasn’t built with LLMs in mind.

There is one tool that all LLMs rely on, yet nobody fully appreciates: search. The world has invested hundreds of billions of dollars to develop the internet search capabilities we now consider routine. We rarely think twice about granting AIs access to a search API—but if this API didn’t already exist, it would require an extraordinary effort for someone to build it from scratch.

If someone creates a critical tool and integrates it with a cutting-edge LLM to achieve valuable tasks, customers will willingly pay a premium for tokens resold by the tool-builder. But these tools aren’t trivial MCP servers executing granular tasks. Instead, they involve large data, complex requirements, high compute needs, distributed systems, and advanced simulations.

A company aiming to develop an agentic engineer, detective, or auditor must first construct these sophisticated tools. Interestingly, once critical mass is achieved, the tool-maker can leverage feedback loops to enhance models continually, establishing a durable competitive advantage. One can begin by using reinforcement learning to optimize their own tools, eventually training models from the ground up.

At this stage—with the tools, the feedback loop, and proprietary models—a company’s market position becomes virtually unassailable. So far, no one has reached this stage. Companies like Cursor mention training their models but still primarily rely on external foundational models like Claude.

The opportunity here is clear: whoever bridges these tool gaps effectively and first builds the comprehensive ecosystem described above will capture extraordinary value for their vertical.