So, you’re building agentic software?

With all the feedback we get, how do we know the difference between signal and noise?

A student of mine asked me that yesterday while I was coaching them on the very first product they’re launching. That question right there is what Design is all about in this age of agentic software.

The product in question is one we’ll be seeing more and more of: an AI-native SaaS that helps small agro-business companies get clarity from their commercial data and make the right decisions (what to buy and sell, at which price, through which provider, when). Reduced to its simplest expression, it’s a dashboard and some reports, paired with 2 AI agents supposed to act as a CRO and CFO - of course.

This MVP was already pretty mature. Clean design patterns. Already the ersatz of a brand. Navigable and understandable. Some tweaks here and there, but very acceptable overall.

Consequently, all visual distractions went poof, and for one hour, we got to discuss what really mattered: information architecture. How to present data? Which information to prioritize? How to go from information to action?

As we were balancing options, my student got overwhelmed: with all the feedback they were receiving from coaches, advisors, and users, how could they distinguish between signal and noise?

There’s no easy answer for this one, but part of the trick is to orient that question differently: since you’re building an agentic software, i.e. a decision-making tool where your users are reviewing, validating, and orchestrating more than doing themselves, your number one priority becomes figuring out the difference between signal and noise for them.

And to do that, you need to know them well. Spend time with them. Talk to them. Ask open questions about their environment, their habits, their challenges, their hopes. Go observe how they work. What vocabulary they use. How they set their priorities. Take notes, pictures, videos (while asking permission, please). Distinguish what your customers do from what they say, and you’ll get the signal instead of the noise.
There are proper tools to do all that, but mainly read Just Enough Research by Erika Hall and you’ll be fine.

Once you do that, go iterate on your information architecture by asking this single question: how can you get your user closer to achieving their goals, instead of merely displaying data to inform their decisions? If you’re serious about agentic services, get serious about understanding what « done » means for your customer.

When is a task truly done for them? Are you able to get them to that point, or are your agents merely chatting their way into a workflow that could have been a form or a stepper?

Good design is all about alleviating cognitive load by figuring out your users’ mental model. You’re not doing that when you confront them with a chatbot that forces on them an exclusive read/write interaction. It’s cognitively exhausting. So bring your design components to your conversational UI. Bring buttons, dropdowns, drawers, forms, cards, thumbnails and all the crew to the chat box.

Standard design already figured out how to address so many cognitive tasks: comparison, selection, review, validation, rating, planning, sharing... It’s time that conversational patterns joined forces with structured components to really unleash the power of agentic software, without exhausting our collective brains.

Who said that a conversation should only be text?