Use this whenSomeone asks for a new app, dashboard, or portal, and you’d like to know if an app is the right answer.

Why it exists

You don’t want the organization spinning up apps. Every new app is another interface, another login, and another thing to train, secure, and maintain. A product can solve a narrow problem and still give the organization more work overall. If AI is supposed to remove work, a new place for people to manage it is an incomplete answer.

The framework

  1. Step 1Capability, not appWhat needs to happen?
  2. Step 2Headless serviceWhat do the data and actions need to be reachable?
  3. Step 3Agent as the interfaceWhere do people already work?
  4. Step 4Interface when earnedDo people need to see, compare, or approve?
  1. Capability, not app

    What needs to happen?

    Say the need as a verb and an outcome, such as “assemble the current project picture and flag what needs a person.” The screen is one possible way to get there.

  2. Headless service

    What do the data and actions need to be reachable?

    Expose the data and the actions behind an API or MCP, with permissions and an audit trail. The service does the work whether or not anyone is looking at a screen.

  3. Agent as the interface

    Where do people already work?

    Let an agent carry the request in the place people already are, such as chat, and return something they can use, including a clear failure when it can’t finish.

  4. Interface when earned

    Do people need to see, compare, or approve?

    Build a screen only for what an agent can’t do well. Keep it as small as the job, and design it for checking and correcting, not for moving information around.

How to use it

  1. When an app request arrives, write the need as a capability: a verb and an outcome.
  2. Ask the question that comes before any build: what would an agent need in order to do this?
  3. List the data and actions involved, and check whether they are already reachable, with permissions, through an API or MCP.
  4. Put the capability where people already work, and let an agent call it.
  5. Add a screen only when the list below says it is earned. Keep it small.

A screen is earned when

  • People need to see many things at once, such as a schedule or a portfolio.
  • People need to compare options side by side before choosing.
  • The work is visual, such as drawings, layouts, or models.
  • Someone has to approve or correct a lot of items, and a list of chat messages would be slower.
  • Oversight needs a place to look, such as a log of what the agent did and why.

In practice

At Rivian we connected systems through custom APIs and MCP, with Slack as a place people could work across them. The idea is the same in small: the capabilities sit behind the APIs, and people meet them in the place they already work.

Where it breaks down

Some work is visual or dense, and an agent in a chat window is the wrong interface for it. Headless only works with real permissions and audit trails, because an agent with broad access and no record is a risk. And people can’t use what they don’t know the agent can do, so someone has to tell them. The agent is the default interface, not the only one.

Try it with AI

Someone has asked for a new app. Help me decide whether we need it.

Ask me one question at a time. First, what needs to happen, as a verb and an outcome? Then ask what an agent would need in order to do this: which data, which actions, which permissions, and which audit trail. Then ask where the people involved already work.

Finally tell me whether a screen is earned, using these tests: people need to see, compare, or approve many things at once, the work is visual, or oversight needs a place to look. If not, describe the headless version.

Related