AI agents are starting to use your app. Is it ready?

For years we built apps for one kind of user: a person, with a screen, clicking buttons. That is changing quickly. More and more of the work inside apps is about to be done by AI agents acting for a person.

This month Google started turning Gemini into an agent that can act across business software. It connects to tools like Google Workspace, Microsoft 365, Slack and Jira, and to any app that offers an MCP server, the new standard way for an AI agent to talk to an app. Google is launching it with businesses first.

At the same time, some of the fastest rising projects on GitHub are about making software easy for agents to drive, through simple commands instead of screens.

The direction is clear. Soon a lot of your users will not open your app at all. They will ask their assistant, and the assistant will use your app for them.

What an agent needs from an app

An agent does not see your beautiful screens. It needs four things.

• A way in that is not a screen. An API or an MCP server lets an agent ask for data and take actions directly. Without one, the agent has to click through your pages like a robot, which is slow and breaks every time the design changes.
• Plain descriptions of what each action does. Agents choose tools by reading their descriptions. “Create invoice: makes a draft invoice for a customer, does not send it” gets used correctly. A vague name does not.
• Clear limits. Which actions are safe to repeat, which ones send an email or charge a card, and which ones can never be undone. Good agent tools say so up front.
• Permissions that match a person. The agent should only be able to do what the person it works for could do, and every action should be logged under that person.

Where to be careful

Giving an agent the keys to your app is powerful, and a little scary. A few rules we follow:

• Drafts before sends. Let agents prepare things such as emails, posts and invoices, and let a person press send. Our own OwnSites MCP tool writes blog posts as drafts only. It never publishes on its own.
• Read first, write later. Start with tools that only look things up. Add tools that change data once you trust how they are used.
• Rate limits and budgets. An agent can repeat an action a thousand times in a minute. Cap it.
• A clear record. Keep a log of what every agent did and when, so mistakes are easy to find and undo.

How we build for this at MashData

Every app we build now gets two front doors: one for people and one for agents. The people door is the screens you know. The agent door is a small, well described API, and for many apps an MCP server, with the same permissions and checks underneath.

It costs little to plan for at the start and a lot to bolt on later, the same way phone support did ten years ago.

What this means for you

If you are planning an app, add one question to your list: what would an AI assistant need to do here for my user? Answer that early and your app will be ready when your customers start asking their assistant instead of opening the app.

If you would like help building an app that works for people and for their agents, tell us your idea.

Sources: TechCrunch: Google brings agentic AI to Gemini, starting with businesses