Utbenat: Agency

Introduction

Sebastian:
Imagine you're cleaning the kitchen. You've vacuumed, and it's time to mop. You take out your trusty mop. The mop is a fantastic tool. It gets the floor sparkling clean, and it's much better than scrubbing on your knees with a rag. But the mop has an obvious limitation: it does nothing until you're holding the handle. You have to decide exactly where to push it, how hard to press and when to change the water. It needs your constant hands-on attention.

Now imagine you buy the very latest robot vacuum, one with a built-in mopping function. You don't stand there holding it. You press a button, or say "Clean the kitchen". The robot draws a map in its head, navigates around the chair legs, doses the water and goes back to its charger on its own when it's done. You gave it a goal, and it worked out the way there by itself.

Today we're going to talk about the difference between the mop and the self-driving cleaning robot. We're going to talk about agency. It's about the next step in AI, where the technology goes from being a passive tool waiting for step-by-step instructions to becoming a more independent actor, with agency of its own, that solves the tasks we built it for. But above all, we're going to talk about what this demands of our systems if it's going to work as well as we want it to.

From prompt to goal

Sebastian:
Agency, in an AI context, means the ability to act independently to reach a goal. Until now, AI has mostly been like the mop. We've given it a specific instruction, "Write a draft of an email", and the AI has written the email. Then it has stopped and waits for further instructions.

An AI agent works differently. You give it an overarching mission. You build it to, for example, "Monitor the support ticketing system and answer the questions covered in our internal FAQ". The agent continuously reads incoming tickets, recognises the most common questions, looks up the right answer in your FAQ, writes a reply to the customer, sends it and closes the ticket. It reasons, plans and carries out the task in several steps, entirely on its own. Fantastic!

The value of agency: scaling time

Sebastian:
The value of this is clear. When AI goes from being an assistant in a chat window to being an agent in your systems, we can free up time for real. We go from AI helping us work faster to AI actually handling entire tasks while we focus on more complex work.

But there's a catch. For agents to do their job, they have to interact with your systems. They need to get into the CRM, into the ticketing system and into the e-commerce platform. And this is where we realise that we've spent the last few decades building systems designed for people, simply because people were the only ones using them.

Systems built for eyes and a mouse pointer

Sebastian:
By and large, we've built our software to be navigated with our eyes and a mouse pointer. We talk about "usability" and UX. It's about nice layouts, clear buttons, forms and drop-down menus. The problem is that AI agents don't have human eyes. To a machine, the visual interface is often just unnecessary noise. An agent doesn't want to click its way through three different submenus; it wants to interact programmatically. It wants to send and receive clean data directly from the system's engine.

So what happens if a system lacks this programmatic access? It's not that the agent absolutely can't solve the task. You can use techniques that help the AI "read" the screen and simulate mouse clicks. But it becomes incredibly fragile.

The system then has to run on a virtual machine, or your laptop, that renders the visual interface, which takes more AI power and makes the process more expensive and slower. Worse, if the vendor ships an update where a button suddenly moves, the whole process risks failing. It simply doesn't work very well to put it into secure, scalable production. If we build our workflows around visual interfaces, we drastically limit the agents' speed and reliability.

Programmatic access and the service entrance

Sebastian:
So what do we do? Well, to make full use of agency, we have to change how we write requirements for and buy software. Beyond asking how user-friendly a system is, we have to start setting requirements for what's called programmatic access. We need to make sure systems are built with an "API-first" mindset. Can a machine understand and interact with this system securely, without having to pretend to be a person clicking on a screen?

Imagine you've built an incredibly beautiful head office. There's elegant stucco work, heavy oak doors with sleek handles and soft lighting. It's a fantastic environment for the employees. But one day you decide to streamline your services and buy small self-driving robots to deliver coffee, lunches and office supplies to the meeting rooms. What happens? Well, the robots can't push down the door handles and open those beautiful doors.

Upgrading the robots' intelligence won't help. You could of course buy much more expensive and more fragile robots, but it's easier to adapt the building. You install automatic door openers. Our IT systems are like that building. We've built beautiful environments for people, but when we want the agents to start working, the systems need their own programmatic ways in.

Three questions to ask on Monday

Sebastian:
Here's what you do on Monday. The next time you're in a meeting to evaluate a new system, a new tool or a new platform, ask the vendor or your own IT department three plain questions:

• Is there a modern and complete API? In other words, is there a dedicated "service entrance" where programs and AI can exchange data directly, without going through the visual interface?
• Is the information stored and exported in open, standardised formats, or is the data locked in proprietary, closed formats that we can't reliably read from and write to?
• Can the work be done fully programmatically, or does the system force us into the visual interface?

If you start asking these questions with every purchase, you'll build, step by step, an infrastructure where your AI agents can work reliably and deliver the value they have the potential to deliver.

Summary

Sebastian:
So, to go back to where we started: right now, across our organisations, we're swapping the manual mop for self-driving robots. But no matter how smart that new AI agent is, the whole point falls apart if it ends up stuck in front of a heavy, closed oak door and can't get into the systems.

We have to make sure those automatic door openers are in place. By requiring open APIs and programmatic access already at the purchasing stage, we make sure the agents can actually do their job. Only then do we go from cool AI experiments on a screen to real business value.

Thanks for listening to today's episode of Utbenat! If you want to talk more about today's topic, or just talk data, get in touch. We're on LinkedIn. Or contact us through our website, fiwe.com, with a double-u. Talk soon, and take care out there!

You've been listening to Syntra, a podcast from Fiwe.