> ## Documentation Index
> Fetch the complete documentation index at: https://docs.anymorph.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Agent
> A chat that reads your workspace's real records — and always asks before it changes anything
**Agent** is a chat window where you can ask anything about your own workspace: visibility, prompts, citations, pages, traffic. It answers from the records in this workspace, not from a guess.
It can also do work — change the prompt panel, edit and publish pages, update your brand documents and generation rules, add knowledge, start an analysis or publish pages. Anything that changes the workspace **stops and asks you first**.
The sidebar shows a **Beta** badge next to **Agent**. It's shipped and usable, and it's still changing.
1. **New chat** — start a fresh conversation.
2. **History** — open your saved conversations.
3. **Starter questions** — click one and it sends immediately.
## What's on the screen
**Toolbar**
* **New chat** — opens an empty conversation. A thread you open but never write in is not saved to History.
* **History** — lists your saved conversations, newest first. Each is named automatically from your first question and the name never changes, so the list stays scannable. An unnamed one shows as **"Untitled chat"**. With no saved conversations the button is greyed out.
**Start screen**
An empty thread shows **"What do you want to know?"** and four starter questions: how visibility moved over the last 30 days, which prompts you lose most often and to whom, which sources AI engines cite about you, and finding your biggest coverage gap. Click one and it sends.
**Conversation**
Your messages sit right-aligned in a grey bubble. The agent's replies are on the left, as formatted text with headings, lists, tables and links. Before anything is written you'll see a shimmering **"Thinking..."**, then a row for each thing the agent is reading, with a short summary of what it found.
**Message box**
* The placeholder reads **"Ask anything about \[your workspace name]..."**.
* Enter sends. Shift+Enter makes a new line.
* While an answer is streaming, the send button becomes a stop square and **"Working..."** appears above the field.
* You can keep typing during an answer. Extra messages queue under **"Sends when this answer finishes"** and send by themselves. Each queued message has **Edit** and **Remove**.
## What it reads, and what it can do
| Category | What that covers |
| --------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Reads without asking | Your visibility numbers and trends, citations, traffic summary, prompts and prompt diagnostics, pages with their content and research, catalog products, brand profile and rules, knowledge search, run status |
| Reads without asking, but costs time or money | Web search and page fetches, keyword data, live search-result checks, Google Search Console performance and URL inspection, competitor-gap lookups. It is instructed to use your own records first |
| Always needs your **Apply** | Prompt panel edits; creating, editing, publishing or rebuilding on-site pages and linking products to them; catalog product changes; brand profile, global rules and design guideline changes; knowledge entries and research sources; starting a visibility analysis or a publish run; generating an image |
| Cannot do at all | Workspace status and domain administration, destructive catalog deletes, indexing work outside this workspace, and running arbitrary code. Those stay with Anymorph operators |
## How approval works
Say it plainly: "re-target the prompt panel for this topic", "publish this page", "update the brand profile".
The agent stops and shows a card headed **"Review proposed changes"**, listing the exact values that will be written, one field per line.
Risky actions carry a red warning above the buttons. Anything that starts a generation run warns that it spends budget and takes minutes; publishing a page says *"Publishes to the live tenant site."*
**Apply** runs it, then the agent reads the record back and reports what actually happened. **Discard** means nothing happens, and it won't re-propose the same change — it'll ask what you'd want instead.
Approving a publish or an analysis starts real work outside the chat that takes minutes. Results appear on [On-site](/dashboard/on-site), [Visibility](/dashboard/visibility) and [Performance](/dashboard/performance). You can ask the agent how a run is going.
## Questions that work well
The agent is instructed to lead with the answer and put the evidence after it, always state the time window as a number, use tables for comparisons, and say plainly when there's no data instead of guessing. It replies in whatever language you write in.
* "How did mention rate move over the last 30 days? Break it down by topic in a table."
* "Which 10 prompts do we lose most often, and which brands get named instead?"
* "Show me the top 10 source domains ChatGPT cites about us."
* "List the on-site pages published last week and whether each one is indexed."
* "Find the topic where we're furthest behind competitors, and suggest what page to write."
* "Update the brand profile so prices are always written in both KRW and USD."
* "Add five prompts about free trials to this group."
* "Which five pages gained the most citations versus last month?"
## Limits
* One answer runs at most **80 internal steps**. A very sprawling request can end before finishing everything, so ask in smaller pieces.
* A question that walks several datasets in sequence can take tens of seconds. A simple lookup is quick.
* It is locked to **the one workspace** you're in. It can't read or change another.
* **History** shows your **30** most recently used conversations.
## Status messages
| What you see | What it means |
| ---------------------------- | --------------------------------------------------------------- |
| **Thinking...** | The turn started, nothing written yet |
| **Working... / \[n] queued** | An answer is streaming; n of your messages are waiting |
| **Reconnecting...** | The connection dropped; it retries automatically |
| Red connection error | Retries ran out. Send another message to pick it back up |
| **Stopped** | You cut the turn short. Not an error — the partial answer stays |
## FAQ
No. Conversations are stored against your own login within the workspace. You can't open theirs and they can't open yours.
No. Every change is proposed as a card and only runs when you press **Apply**.
The card lists each field that will be written. Anything that spends budget or publishes to your live site carries an explicit red warning line.
Nothing happens. The agent moves on and won't re-propose the same change; it'll ask what you want different instead.
Yes. The run continues, and reopening that conversation rejoins it in progress.
Yes. Your message queues and sends automatically when the answer finishes. Until then you can **Edit** or **Remove** it.
Yes. It replies in whatever language you write in.
No. It's limited to the workspace you're in.
## Related
The workspace at a glance. Agent is where you ask it anything.
How to change the prompt panel from chat, and what to watch for.
The brand documents the agent reads and can update.
What to look for when a draft lands.