Answers you can interact with

When an answer is better seen than read, the agent draws tables, charts, sliders, maps, quizzes and math right in the chat.

An interactive answer is a part of the agent’s reply drawn as live UI: a table, a chart, sliders you drag, a map, a quiz. The rest of the reply stays ordinary text, and the UI builds as the agent writes. To put a chart or a form in your own app instead, ask the agent to build it and open it in the Preview. For a slide deck, see Slides, drawings and diagrams.

Alpha will:

1. Answer with the balance at your numbers, in a sentence.
2. Draw sliders for the deposit, the rate and the years, beside the balance and a chart.
3. Recompute every figure in your browser as you move a slider. No new turn starts, so it costs nothing.

Show me, or build me?

A question about information is answered in the chat, even when the whole answer is charts. Something for your app is code in your project, and nothing is drawn.

Show me: answered in the chat	Build me: code in your project
”Compare Postgres, MySQL and SQLite for a small SaaS"	"Add a pricing page"
"Show me signups by week"	"Build a dashboard of signups for my app"
"What do I have after 20 years at 7%?"	"Make me a mortgage calculator"
"What are the best day hikes near Seattle? Show them on a map."	"Build me a store locator page with a map for my app”

“How does HTTPS work?” has nothing to chart, so it stays text. The figures in a block are always real: computed, read from your files or database, or given by you. The agent never draws placeholder numbers.

A request that reads either way can land on the wrong side. Say “in the chat” or “build it into the app” to settle it.

What can the agent draw?

You don’t name the block. Ask the question and the agent picks one.

Block	When it’s used	A prompt to try
Tables and comparison cards	Options to weigh side by side	”Compare Postgres, MySQL and SQLite for a small SaaS”
Charts and figure tiles	Numbers over time, a breakdown, a total	”Show me signups by week”
Buttons and short forms	A choice, or two or three details the agent needs before it goes on	Appear on their own when the agent needs your input
What-if sliders	Figures that turn on numbers you might change	”What do I have after 20 years at 7%?”
Maps with numbered pins	Where the places are is the point	”What are the best day hikes near Seattle? Show them on a map.”
Checklists with step timers	Something you tick off as you go, such as a recipe’s timed steps	”I’m hosting a Sunday lamb roast and don’t know how many are coming yet”
Steppers	A count the figures follow, such as guests or servings	The same lamb roast prompt
Simulations you can run again	Chance that an explanation rests on	”Help me understand the Monty Hall problem”
Quizzes	Testing what you’ve learned	”Quiz me on the solar system”
Flashcards	Studying a set of facts	”Make me flashcards on the parts of a sailboat”
Photos	A place, dish, animal or object worth seeing	”Show me the five most iconic dishes of Rome, with photos”

Blocks can also sit in tabs, accordions and carousels. Photos come from Wikimedia Commons, credited to their author and licence. A trip plan stays a list: the agent draws a map only when where things are is the point.

Talk back through an answer

Some controls work only in your browser. Others send a message, which appears in the transcript as your own, so nothing a click sends is hidden.

Control	What happens	Reaches the agent?
What-if sliders and their fields	Every figure recomputes in your browser. Values settle when the pointer pauses, so a long drag updates in steps.	No
Buttons, list items and follow-ups	Their text is sent as your next message.	Yes
Forms	Submitting sends the button’s text plus one name: value line per field you filled in.	Yes
A clickable card	Sends its text plus a Selected item: {…} line with the card’s details.	Yes
Links and copy buttons	Links open web pages (http and https only) in a new tab. Copy buttons copy to your clipboard.	No
Quiz	Each answer is graded on the spot. Next moves on, See results shows how many you got right, and Try again starts over.	Only Go over what I missed, which sends your misses
Flashcards	”Tap to flip”, then Didn’t know or Knew it. At the end, Review the … you missed or Start over.	No
Timer and stepper	The timer’s start, pause and reset buttons, and the stepper’s − and +.	No
Map	Tap a pin for its note. The scroll wheel or one finger scrolls the chat; ⌘ with the wheel, or two fingers, moves the map. Use Ctrl instead of ⌘ on Windows and Linux.	No

A block’s buttons stay disabled until the agent’s text moves past that block, not until the whole reply ends. A click sent while the agent is working always steers the running turn, whatever your default for typed messages. If the turn can’t take it, the message waits in the queue instead.

Read formulas in a reply

TeX in a reply or in the agent’s reasoning draws as typeset math, inline or centred on its own line. Codex’s \(…\) and \[…\] and Claude Code’s $…$ and $$…$$ draw the same.

Explain gradient descent for linear regression, with the math.

Prices survive next to formulas: $5 and $10, $HOME/$USER and ($$) stay as written. A formula shows as source until it is closed, and one wider than the chat scrolls sideways.

What are the limits?

Limit	What it means for you
State isn’t kept across a reload	Slider settings, form values, ticks, timers and quiz progress reset. What a click already sent stays in your message.
No live data	A block shows the figures the agent had when it wrote. Ask again for fresh ones.
Map lines are straight	Lines join the stops directly, not along roads. Pins are only as exact as the coordinates the agent found.
Photos need the internet	With no network on the session’s machine, or no good match, the agent draws without them.
A block that fails shows its code	You see its source as a code block, so the answer is never blank. A formula that can’t be parsed shows in red.
Viewers look but can’t send	People you share the session with see every block. Clicks that would send a message do nothing for them.
Math draws only in the chat	The Plan tab, the Agent tab and Markdown file previews show TeX as source.

Turn it off

Interactive answers come from Generative UI, a skill that ships with Alpha, on by default for both agents. Switching it off changes only your sessions.

1. Open the Skills page

   Click Skills in the sidebar, then the Installed tab. Generative UI is listed under Your skills.

2. Switch it off

   Use the switch on its row, or open it and turn off Installed in your sessions. The dialog reads “Off keeps it out of your sessions alone.”

3. Send your next message

   Replies are plain Markdown from then on. Blocks already in your transcripts keep drawing.

To ask for a drawn answer outright while the skill is on, mention it in your message: $generative-ui in Codex, /generative-ui in Claude Code. Math draws either way; it isn’t part of the skill.

Admins only Two more switches in the same dialog decide for every account. Ships with Alpha off moves Generative UI to the catalog, so it is only in the sessions of accounts that install it. Let the agent use it on its own off means only a mention runs it.

Next steps

Queue, steer or stop

Choose whether a message typed mid-turn joins the running turn or waits for the next.

Prompt Alpha well

Phrase a request so it lands as an answer in the chat or as code in your app.

Skills

Install or write skills that teach the agent other ways to work and answer.

Follow the agent's work

Everything else a reply can hold: commands, edits, plans and errors.