# Part 10: Showing what it knows | Team0 > Define a page your agent builds from the brain when someone asks: every line sourced, mistakes corrected where they are seen, in any agent. Agent-readable version of https://team0.ai/learn/showing-what-it-knows. The page for people is at that address; this file carries the same content without layout. --- [Building an agentic brain](https://team0.ai/learn) Part 10 # Showing what it knows Define a page your agent builds from the brain when someone asks: every line sourced, mistakes corrected where they are seen, in any agent. Yogev Ben-Tov ยท 5 min read Your agents can now read the brain. People still need to see it: where an account stands, what they owe, what changed while they were away. The usual answer is another dashboard, which someone has to build, maintain, and remember to open. Most of the working day now runs through agents. So don't build the interface. Define it, and let the agent build it from the brain when someone asks. By the end of this part, you will have one app: a page your agent fills from a read, where every line opens its source and a correction made on the page goes back into the brain. ## Define it, don't build it An app here is two files your agent loads as a skill: - **A page template.** One static HTML file that renders entirely from a single data block. It ships with invented example data, so you can open it and judge the design before it touches anything real. - **Instructions.** Which reads to make, the rules for turning lines into fields, length limits for each field, and what to say after the page is shown. When someone asks the question, the agent reads the brain, fills the data block and shows the page, in the chat they are already in. You still own the experience: the layout, the words and the rules are yours. The agent supplies only the data, and only from what the read returned. ## Why let the agent build it - **It appears where people already work.** Any agent that can load instructions and read the brain can build it. Nobody has to go to another tool. - **It is built for this question, now.** The page uses this morning's read, so it is current when it opens. - **It shows only what the reader may see.** The read already enforces access (part 8), so the page cannot show more than the person is allowed to know. - **It can be checked.** Every line keeps the reference from its read, so a click opens who said it, where and when. Uncertain lines stay marked, and a changed value shows what it replaced. - **It teaches the brain.** Corrections made on the page go back through your write path (part 3), so every agent reads the fix from then on. - **It is cheap to change.** Editing a template is not a release. A team can copy an app and change it to fit how they work. It works with any brain. The contract is the read: lines with a stable reference, a trust level and earlier values. MD files, a database or a product can all supply that. ## Where interfaces are going For years a product team shipped one screen, and everyone opened that screen. When people work mainly through agents, the agent is where the screen appears, and it can be built for the question in front of it. What the team owns moves up a level: the data shape, the template and the rules. The experience stays designed, consistent and on brand, and the agent does the assembly. The brain is what makes this safe. Without one, an agent-built page is only as good as what the agent remembers from the chat. ## Know what it is not The page is a snapshot from the moment of the read. It will not update by itself, and asking again builds a fresh one. Each build costs reads and tokens, so measure that the same way as part 9. Setup and administration still need a real application: connecting sources, permissions, inviting people and billing. Keep that application. The page is for seeing and correcting. A filled page is a file containing real information. Decide who may receive it before anyone shares one. ## Build one 1. **Pick a question people ask every week.** Use one from your part 0 set: "Where do we stand with this customer?", "What do I owe?", "Prep me for this meeting". 2. **Write the data shape first.** List the fields the answer needs, their limits, and what happens when a field is empty. A missing field should leave its section out, not show a placeholder. 3. **Write the template.** One file, one data block between clear markers, no network calls. Put the parts every app shares in one place: the source drawer, the "Not right" control and the corrections list. 4. **Write the rules.** Only what the read returned: never a person, date or number from memory. Keep a colleague's statement as a statement, not their words. Never show how the brain works: no read IDs, timeouts or partial-read notes. 5. **Close the loop.** The page collects corrections into a short list with the references attached. The person pastes it to the agent, which saves each item through the write path and reports what was saved. Ask your agent to write the first version from the question and your data shape. Then change the design yourself. ## Measure Start with the invented data. Check the page on a phone, in dark mode and with empty sections. Then run it on real reads for your part 0 tasks, and compare it with the plain chat answer to the same question: - **Faithful:** every line traces to a line in the read. Count invented lines; the target is zero. - **Useful:** the person finds the answer faster, and spots what changed or is uncertain. - **Closed:** corrections arrive in the brain, and the next read reflects them. - **Consistent:** two different agents build the same page from the same read. - **Cost:** time to build, and tokens per page. If two agents fill the same read differently, tighten the instructions or the data shape before changing the design. Team0's own apps follow this pattern; you can [try them on an invented company](https://team0.ai/apps). --- *Build and measure your own setup with the companion skill. [See the series](https://team0.ai/learn). For a maintained company install, [talk to Team0](https://calendar.app.google/uM9ptyv83Ubgnq226).*