Meeting Agents Halfway
I'll describe an idea that's been in my head. One can imagine an AI to interact with all the business software out there.
How? Right now, MCP is an answer. But I suggest a different way that adapts to the existing UX.
Let's take Lattice (an HR and performance management software) as an example
If Lattice were to build an MCP, the developers have to write a manifest file which is basically exposing selective backend functions (mostly reads and few writes). That doesn't solve it yet. The AI still has to think through how to piece those functions together that will lead to running a review cycle for 1000 employees with all kinds of org Structures.
The alternative I'm proposing is through computer use (think operator) but the thinking resides on Lattice's end. Hear me out.
I had a conversation with the Lattice CPO and he was asking for an agent action to sit on Lattice app, the purpose of it is to run the performance review cycle end to end (with history of how the founder has ran it). Now, what's interesting is this ai agent action will know the history of the user actions and will Lattice certified in terms of how to maneuver the product and get to the result.
So what I'm proposing is a B2B SaaS idea where like intercom that sits on a web app, we will have something that is there for Computer Use agents to ask us what they want to get done. By that I mean Computer Use but the thinking of how the app works is residing in the business software end.
The reason i'm trying to emphasise the thinking to reside on the Lattice end is because: What is Lattice? they are the world experts at running large scale performance reviews.
They used to translate that expertise using a UX with a web app etc.. Now there can be a new way: Either MCP exposing functions to run the whole workflow or an agent running the flow on their web app. I think the latter is meeting the Agent halfway: You made it to my website why don't you run these battle tested flows?
As an entry point, it could look like this: A set of cursor guides inside the app. Where when the user's agent clicks on a guide (like 'Run end to end performance review'), a second cursor shows up and navigates them through the app to run this workflow.
Finally to be concise: If computer use were to do it, it'll be clicking through the app and going back correcting itself. Instead a certified cursor flow that tells exactly where to click is what i'm thinking.