Agents can now search the web

Welcome back to Building AI Agents, your biweekly guide to everything new in the field of agentic AI!

Remember, the difference between a successful product and an unsuccessful one is sometimes just a catchy name. Welcome back to Building AI Agents Model Context Newsletter (MCN).

In today’s issue…

• Anthropic gives its LLMs internet access

• IBM goes all-in on agents

• Y Combinator’s ideas for agent startups

• A course on building voice agents

• Boring business + AI agents = $$$ ?

…and more

🔍 SPOTLIGHT

AI agents just gained a powerful new tool.

Last Wednesday, Anthropic introduced a web search feature to its API, allowing its Claude family of large language models (LLMs) to augment their existing knowledge by searching for relevant information on the internet. When Claude is equipped with the tool, it determines whether a web search could provide useful context, formulates a search query, and uses the results to generate a response, complete with citations to the web pages it accessed.

Not only can Claude intelligently determine when to use the web search tool, but it can also iterate through multiple rounds of searches, agentically refining its query and analysis until the necessary information is found.

This new capability does not come out of the blue—in March, Anthropic added web browsing to its Claude chat assistant, presumably powered by an earlier version of this feature, before the company was ready to roll it out to its API. In both cases, Anthropic is chasing its rival OpenAI, which released web search to its ChatGPT assistant in October of last year, and to its API in March. Google’s Gemini, the third major horse in the AI race, introduced a similar Deep Research capability in December. With Anthropic’s announcement, users of all three leading LLM providers can now integrate the full knowledge of the surfable web into their applications.

But why would all of this be necessary? Aren’t LLMs already trained on essentially the entire publicly-available text of the internet? Why do they need to search something they’ve already “seen?”

Two reasons. First, LLMs are trained on much of the internet, but not all of it. The process of cleaning and filtering data for training removes significant amounts of material—not to mention that some pages rapidly go out of date. Ask an LLM what stadium the Chicago Bears play at, and you will likely get the right answer. Ask it who their next opponent is, and you almost certainly won’t.^1,2

Second, LLMs are effectively compressed representations of the data they’ve been trained on, and compression usually comes with a tradeoff in accuracy. Facts an LLM has seen numerous of times in its training set—like the name of France’s capital—will be recalled accurately, while ones which appear rarely—like the name of the mayor of Paris’ 12th arrondissement—may not. Like humans, LLMs learn through repetition, and looking up information is much easier then memorizing humanity’s entire corpus of knowledge.

So answering questions via web browsing, one of the most basic but fundamental skills in any human worker’s toolbox, is now in the hands of AI agents. Hard-coded, unreliable web scrapers are no longer necessary to build agentic tools that can retrieve nearly any piece of public information in real-time. The applications are limited only by builders’ imaginations.

¹Ask it when the last time they won a Superbowl was, and the answer will probably remain fresh for a depressingly long time.

²A major source of annoyance for me in trying to automate agent building is that—ironically—LLMs are lousy at writing code for LLM APIs, since they change so rapidly.

🤖 AI AGENTS FOR YOUR BUSINESS

Orynt AI: The official implementation agency of Building AI Agents

• 🕙💰 LIMITED TIME ONLY: Refer a future customer to us and get 10% of the total value of the contract as a bonus!

• Discover exactly where AI can streamline your operations, saving you time, reducing overhead, and improving consistency.

• Automate routine tasks with agents tailored specifically to your business needs.

• Unlock serious growth and efficiency by scheduling your free 30-minute strategy call.

• See a real-world client AI agent build here