How AI agents got good at using tools

Anthropic's MCP could do for agents what TCP/IP did for networking.

It’s day two of Agent Week! This post is for subscribers only, so if you aren’t yet a paying subscriber I encourage you to click here and get 20 percent off an annual subscription.

---

When ChatGPT was first released in 2022, it was missing something important: the ability to interact with the outside world.

• It could give generic travel advice but had no way to search for available flights or hotels.

• It was surprisingly good at writing and debugging computer code, but users had to manually copy code into the chat window—and then copy the results out.

• It had no way to access workplace platforms like Google Docs, Slack, Notion, or Asana.

In a 2023 article, I argued that this was a strategic opportunity for OpenAI: if OpenAI invented a standard way for LLMs to communicate with the broader Internet, it could cement ChatGPT’s dominance of chatbots in much the same way the App Store has bolstered the long-term success of the iPhone. But that’s not what happened.

In March 2023, days after the release of GPT-4, OpenAI announced plugins. These enabled users to use services like Expedia, OpenTable, and Instacart from within ChatGPT. But plugins didn’t get much use.

Then in the fall of 2023, OpenAI replaced plugins with GPTs. These were custom chatbots optimized for specific purposes. A GPT could have a capability called an “action” that enabled it to communicate with a third-party service. Instead of accessing Expedia from regular ChatGPT, users might use a special-purpose travel chatbot with the capacity to talk to Expedia and other travel-related services. But GPTs didn’t really catch on with consumers either.

Anthropic has taken a different approach. Instead of focusing on its consumer chatbot, Anthropic has prioritized tool-using agents for business applications. This strategy started to gain momentum after the June 2024 release of Claude 3.5 Sonnet. As I wrote last month, this model enabled coding platforms like Bolt.new, Loveable, and Cursor to get traction.

Then in November, Anthropic announced Model Context Protocol, which connects models to external tools. MCP became an industry standard within a few months: OpenAI adopted it in March, and Google followed suit in April.

The combination of long-reasoning models and the open MCP standard means that the industry is entering an age of powerful AI agents. The next generation of AI systems won’t just be able to answer abstract questions. They’ll be able to look up specific information about you or your company and take actions on your behalf. And they’ll be able to solve problems that might take a human worker minutes, hours or perhaps even days.

This transformation has already begun for some computer programmers, who increasingly spend their time reviewing code written by AI systems rather than writing code themselves. I expect other professions to start seeing similar trends in the coming years—perhaps even months.

In this article I want to explain the basics of tool-using AIs. I’ll explain the simple mechanism LLMs use to invoke external tools and why tools and long-term reasoning go together like peanut butter and jelly. Then I’ll discuss how the rapid rise of MCP in recent months could make AI agents dramatically more capable.

How LLMs use tools