Skip to content
The AI Agent Factory

20.16 How the pieces fit

Approved by
Panaversity ·
5 min read

In everyday life. A shop has goods, staff who know how to sell them, a shop window, and a door with a lock. Each answers a different question, and the shop needs all four.

What you will learn. How the chapter's extensions fit one connector, and the mistakes builders make most.

Why it matters. Each extension answers a different question. Mixing them up leads to the mistakes below, such as treating a screen as a permission.

What does each layer answer?

QuestionAnswerBrightline example
What can the AI do?Tools, plain MCPlist_open_invoices, find_vendor
How should the AI do the job well?Skills over MCPinvoice-triage
What does the person see and touch?MCP Apps, with A2UI if the AI shapes the screenThe open-invoices screen
Who may use it?Sign-in, with Enterprise-Managed Authorization and Client CredentialsLuis, Cedar Ridge's employees, the report job

One connector, four layers. A large box labeled sign-in, who may use it, holds three layers stacked. At the top, screens, what the person sees and touches. In the middle, skills, how the AI should do the job well. At the bottom, in gold, tools, what the AI can do and what is allowed. On the right, under three ways in, three boxes with arrows into the sign-in box. A person who approved. An employee, through their company's identity provider. A job with its own identity. Below, tools and sign-in decide what is allowed, and skills and screens make the work easier and never grant anything.

Figure 20.16. One connector, four layers. Tools and sign-in decide what is allowed. Skills and screens never do.

What are the common mistakes?

MistakeWhy it hurtsBetter
A screen with no text answerAI apps without screens show nothing usefulEvery tool returns useful text (20.9)
A screen for every answerThe chat fills with panelsShow a screen when the person will look or act (20.2)
Treating a hidden tool or button as protectedAnyone with a token can call any toolCheck every call in the connector (20.5)
Depending on streamed inputSome AI apps never send itWork from the complete input (20.3)
Using window.openai extras everywhereThe screen works only in ChatGPTBuild to the standard, and check before using an extra (20.9)
Letting the AI write the pageLayout, quality and safety varyApproved parts, such as A2UI (20.10, 20.11)
Generic skill namesSkills from different servers collideDistinct names (20.13)
Changing skill files oftenEvery change means approving againBatch and version changes (20.13)
Linking accounts by email aloneA reused address can open someone else's recordsIssuer and subject (20.14)
A job that borrows a person's accessIt acts as that personIts own identity, with the least scope (20.15)
Testing only that the screen appearsA screen can appear and still lose the person's workTest the whole path: sign in, act, reopen, retry
Testing in one AI appSupport differs feature by featureTest every AI app and plan you promise, phones included

The operating rule: let tools and sign-in decide what is allowed, let skills and screens make the work easier, and test each layer in every AI app you sell into.

Check yourself

Question 1 / 3

0 answered

Which layers decide what is allowed?

Owner
Panaversity