Agents Are Better When They Are Opinionated

Previous post | All posts in series | Next post

Recently, I was talking to an AI entrepreneur about his paused start-up. His opinion was essentially that the software industry was doomed: generative AI from the frontier labs is going to remove the need for any other software. While extreme, he’s not alone in this opinion. When founders, partners, and consultants hear Satya Nadella’s “Business applications will all collapse,” they ask me, “What software is left for me to build?” As AI gets more powerful and can author more difficult software, what is the end game for software companies?

I’ve written before that one solution is to author agents. But even here: there are companies (such as Microsoft and OpenAI) that are building so-called “horizontal” agents, so they can accomplish any task without specific programming. If one of these companies succeeds, authoring agents now just delays the inevitable.

But opinionated agents can never be replaced by general purpose, horizontal agents. Let’s talk about why.

Opinions and Preferences

From dry cleaners to management consulting firms, we have a lot of options when we want to hire someone to do a job. Do you have a preferred landscaping company and a regular Mexican restaurant? Or when buying a product—how many ice cream shops are near you? I bet you can taste a difference and have a favorite, even if they all serve combinations of frozen cream, milk, and sugar.

Large language models produce average content. Without specific instructions, you’ll get the most common possible answer. That’s just how they are trained; models output text based on the most frequent input text.

There’s a parallel with foundational software companies. Products like Microsoft 365 or Google Workspace aim to serve as many users as possible. Giant software products do not express opinions about what your purpose should be when using their software, just like LLMs giving you a middle-of-the-road answer. They are designed to be acceptable to everyone.

Average answers can be OK in the same way that every ice cream is enjoyable. That doesn’t stop us from wanting the one we like best!

As producers of any good or service, we use our taste to create a product that we believe our market will love. The best software does this too.

Opinionated Software

Software that takes less-common approaches is getting more popular. Take Napkin, a diagramming and presentation tool. Diagrams created from Napkin have a very specific look:

All Napkin diagrams are uniformly weighted outlines in black. Because there are scenarios where this style doesn’t work, AI won’t produce diagrams like this by default.

I asked several image generation models to produce, “Simple word art diagram of one light ray entering a prism, splitting into two rays”

Each model attempts, and fails, to produce a scientific diagram—not a word art diagram.

How to Make Opinionated Agents

However, you can emulate the Napkin style with careful prompting. This is Microsoft Designer with an opinionated prompt:

A black outline of a simple triangular prism with uniform weighted black lines, set against a plain white background. The prism is depicted in a clean, minimalist style, with sharp, precise edges and a geometric, symmetrical shape. The lines are consistent in thickness. Designed for simple iconography to be used in an infographic.

This is much better! Writing should be opinionated too. I’ve written before how much I despise text written from magic wand buttons; no one wants to read the extremely average output it creates.

Here’s a very opinionated prompt:

Develop the brief as if it were a manifesto. The language should be dense, provocative, and borderline poetic, designed to make the reader pause and rethink their assumptions. Structure the report into philosophical pillars—Innovation, Authenticity, Vision. Break the rules of conventional business writing by weaving in cultural references, unexpected analogies, and questions that feel existential. This isn’t a brief; it’s a statement.

I had Copilot help write this prompt, of course! Language models are great at brainstorming prompts for themselves. And here’s Microsoft Copilot with Think Deeper writing a product brief about a bib made from bamboo using that prompt:

Bambib is not merely a baby bib—it is a subversive whisper against the banality of disposable infant care. Imagine a world where the sacred ritual of nurturing isn’t marred by synthetic detachment. Bambib’s bamboo fiber defies the tyranny of plastic, emerging as the harbinger of an eco-conscious revolution. It dares to ask: What if every tear, every joyful drool, could be met with the resilient grace of nature’s own fibers?

You don’t have to like that—but it isn’t generic! Maybe marketing copy is too easy of an example. Can opinionated Agents help with more mundane work?

Opinionated Agents at Work

Imagine you are in Accounts Payable. It’s 4:55 PM on a Tuesday, and you get a call from your CEO. You know they’ve been working to close a deal to acquire a related company. “It’s urgent,” they say, and need you to wire $45 million before a competitor hears about the merger and offers a better deal. You have your AI agent begin the process…

And your agent stops you. This agent wasn’t only designed to help with the menial work of updating account balances; it includes instructions to watch for unusual transfers. Because tt isn’t subject to emotion and manipulation, it stops this scam from succeeding.

This is one of the great things about agents that make them better than automations. They can still automate menial work, but the “intelligence” part of artificial intelligence allows them to help you make better decisions.

Let’s consider some more work scenarios with opinionated AI agents:

Brag about your team’s accomplishments with a focus on the types of work your boss consistently praises, in language that describes them as critical to the company’s future
Roleplay as a reluctant customer that you are trying to close a deal with
Give you critical feedback on your ideas before you pitch them to your colleagues
Summarize your recent email with a focus on the people and topics that you consider important
Apply a distinct visual identity to your slides and videos so they are recognized and remembered

One process that I recommend you try is to get advice about a pending decision. Create several agents: an optimist, a pessimist, a scientist, an economist, a historian, a marketer, and a psychologist. Ask each of them to weigh in. You’ll get much better results than if you just asked for an average response!

The Trick

Opinionated agents are better than AI that gives you the default and average answer. The answers are better, but they’re also more engaging. In the same way that you’ll always choose your favorite ice cream shop, users will choose your agent when its answers are aligned with their preferences.

Fortunately, it’s simple to make sure your agents are opinionated. Consider what role you need them to play, add what considerations a talented human in the role would apply, apply your own preferences and taste, and write those into the agent instructions.

I’m planning one more article in this series, and it’s the most important one. Be sure to subscribe (LinkedIn, RSS) if you haven’t yet!