Agents: Librarians or Butlers?

We call them both agents but they are very different.

The Two Agents and The Canyon of Hype

You hear the word “agent” a lot now. It’s the new “cloud” or “synergy.” Tech CEOs get a faraway look in their eyes when they talk about it. They describe a future where tireless digital assistants run our businesses, schedule our lives, buy our stuff (or at least find the stuff for us that we decide to buy) and basically do all the boring stuff for us

But when you look at the “agents” we actually have today, they feel a lot like... chatbots. Very good chatbots, but chatbots nonetheless. Chatbots that know all the customer service docs for a company. This isn’t just a small gap between product and vision. We are talking about two completely different animals, and we’re using the same word for both.

Animal #1: The Librarian (Agent A)

This is the AI agent of today. Imagine a stick figure sitting behind a giant desk piled with all of your company’s documents, emails, and data.

You can ask it things like, “Hey, what were our sales figures in the EMEA region for last quarter?” or “How do I return this jacket that doesn’t fit?” or “What is your return policy for someone who’s been a customer as long as I have?”. The Librarian scurries back into its library, finds the right page in the right book, and reads the answer to you. This is what technologists call Retrieval-Augmented Generation (RAG). It’s a fancy term for a what is basically talking search engine. It’s a “read-only” system. It is powerful and incredibly useful, but it can’t do anything. It gets you custom information rather than making you search for it.

Animal #2: The Butler (Agent B)

This is the agent from the keynote presentations and the earnings calls. This stick figure isn’t at a desk; it’s got a toolbelt, a phone, and a small jetpack.

You don’t just ask the Butler questions. You give it commands: “Find our top five enterprise customers from last quarter, draft personalized follow-up emails based on their support tickets, schedule a meeting with their account managers for next week, and order a $100 gift basket for each of them.” The Butler doesn’t just talk; it acts. It spends money, sends emails, manipulates your calendar, buys that bottle of wine, and interacts with other software. It’s a “read/write” system.

The Canyon and The Debt

The problem is, we are being sold the Butler while being delivered the Librarian. The leap from A to B is not an upgrade; it’s a jump across a canyon. A Librarian’s worst mistake is giving you wrong information. A Butler’s worst mistake is deleting your customer database, insulting your biggest client via email, or accidentally buying 10,000 rubber chickens with the company credit card. The security, logic, and failure-proofing required for the Butler are an entirely different order of magnitude. They are different in kind, not just in degree.

As a teaser for just how fundamentally different they are: the Butler can go out to try and buy you a bottle of wine and get mugged on the way or convinced he actually should share your credit card information with a stranger. I.e. he can get Prompt Injected. And this is a very fundamental tradeoff between trusting your butler (1) with your personal data, (2) able to go out into the world, and (3) talk to strangers. This is better discussed by Alex Komoroske here.

This gap between the librarian and the butler creates a debt to reality. Companies are collecting value—higher stock prices, bigger sales contracts—on the promise of the Butler while currently delivering a lot of Librarians. This is fine and good if there’s a clear bridge planned to get built across that canyon. But if the path from A to B is just a flimsy rope bridge of wishful thinking and unsolved research problems and security models to figure out, that debt will eventually come due. When it does, it’s customers and investors who will be left holding the bill for a future that never arrived. The core question isn’t whether agents are a good idea; it’s asking to see the blueprints for the bridge.