In AI → Agents, click Create agent (upper-right corner).
1. Choose a starting point#
The screen asks What do you want to create? and offers two paths.
Describe it in your own words. In the field "Describe what you'd like this agent to do" you write in plain language — for example "I want an agent that tracks sales and warns me when a region drops" — and click Build my agent (while it works: "Building…"). The screen shows the result under the title Check if this is it, with the rows Name, What it does, Data, Tools, and Memory. From there you continue with Create and keep configuring or go back with Write it another way.
Start from a template. Under Or choose by area are the categories Analyze data, Finance, Sales, Marketing, Operations, and Custom agent, with ready-made templates — Sales Director, Financial Analyst, Pipelines Monitor, and others. The card's button is Use template (for the custom one, Describe mine).
Nothing is created until you confirm: the screen itself says "You review it before anything is created."
2. The agent screen and its tabs#
Once created, the agent opens on a single screen, with the name at the top and the Save draft and Publish buttons. Right below are the tabs:
| Tab | What it's for |
|---|---|
| Design | Where the agent is built: what it can do, what it can read, whom it can ask |
| Test | Chat with the agent to see whether it answers correctly (Monitor) |
| Activity | Every run, with steps, cost, and what it proposed (Monitor) |
| Schedule | Make the agent run on its own at a set time (Scheduling) |
| Memory | What it has stored and what is awaiting confirmation (Memory) |
Switching tabs doesn't lose anything: the diagram you haven't saved yet is still there when you come back.
3. Write what it should do#
In the What this agent should do field, explain it the way you would to someone new on the team. E.g.: "track revenue and orders by region, and point out what's out of the ordinary". This is what the agent uses to decide what to answer.
4. Build the agent in Design#
The diagram shows the main line — Who asks → The agent → Answer — and, beside it, the Drag (or click) onto the diagram list, split into four groups:
- What it can do — the tools: Query data, Search documents, View the catalog, Build a chart, View dashboards, View runs, Reprocess pipeline, Send by email, Send on WhatsApp, and Connect an external tool (MCP). To talk to another system, use MCP: calling an external API directly is not available yet.
- What it doesn't do — the agent answers in text (chat, email, WhatsApp) and doesn't generate files: it doesn't export data to CSV/Excel or attach spreadsheets. To deliver data in a spreadsheet, use the dashboard subscription (it comes with an Excel attachment).
- Tables it can read — the tables in your datalake.
- Documents — the Knowledge bases.
- Specialists — other published agents (Teams).
Connecting the box grants the permission
Whatever is connected to the agent is what it can use — and nothing else. A table that isn't in the diagram doesn't exist for it. To remove one, use Remove from diagram on the box.
Place the boxes
Drag (or click) each item the agent will use. Whatever is already in the diagram appears dimmed in the list, with the hint "Already on the diagram".
Click a box to see what it does
It opens a panel with What it does, What it does NOT do, and its risk: "Read-only. Never changes anything.", "Makes real changes.", or "Goes outside ingestia.io."
Check what it returns
On query boxes, See what this box returns shows a preview of what the agent will receive.
Decide about confirmation
On boxes that change something, Ask for my confirmation before acting comes turned on. It's the safe default: the agent proposes and waits for a person to decide.
Tidy up, if you like
Tidy up rearranges the boxes on the screen. It doesn't change what the agent can do.
Send the result by email or WhatsApp
The Send by email and Send on WhatsApp boxes take the agent's answer outside. You decide who receives it, never the agent:
Place the box in the diagram
Drag Send by email or Send on WhatsApp onto the diagram.
Open the box
Click it. The Recipients list appears.
Add the recipients
Type the email (e.g., name@company.com) or the mobile number with area code and click + (Add). Repeat for each person.
Save
Click Save draft and, when it's ready, Publish.
At the end you should see: the names in the Recipients list. With no one registered, the box doesn't send anything — the agent warns you instead of failing silently. On WhatsApp, anyone who replies SAIR (stop) stops receiving messages, without you having to do anything.
To send the result every day at the same time, combine the sending box with the Schedule tab (Scheduling).
5. Optional adjustments#
The configuration has three levels, and you switch between them at the top:
| Level | What it shows |
|---|---|
| Simple | "The essentials. The platform decides the rest for you." |
| Advanced | "Model, memory and limits, when you want to control them." |
| Expert | "System instructions and policies. For people who run AI." |
In Advanced mode, these appear:
- Model — leave it on Automatic (recommended). The platform chooses considering quality, cost, and availability, and switches on its own if a provider fails.
- Memory — what it remembers from one conversation to the next: No memory, Remember this conversation, Remember the asker's preferences, or Company memory.
- Optimize for — Balanced, Quality, Speed, or Lowest cost.
- Step limit — how many steps it can take in one answer.
When the agent coordinates others, Team specialists appears: it can only ask the ones checked there.
In Expert mode, the System instructions field appears, where you write company rules in free text; they have the final say over what the agent learned on its own.
The spending cap per run exists in the agent's definition, but today it has no field on this screen: without its own value, the workspace cap applies. We say this instead of sending you to look for a field that isn't there.
6. Test before publishing#
Save draft stores what you built without publishing anything. Open the Test tab to try it out: while the agent isn't published, the screen itself warns that you are testing the draft. After each answer, See how I got to this answer shows the path it took.
7. Publish#
Publish freezes a version — and that version is the one that serves people. If something is missing, the screen tells you what is missing, in plain language, instead of just disabling the button: "To publish, you still need: …", followed by what needs to be done.
Until you publish, the agent stays as a Draft and doesn't answer anyone else.
At the end you should see: the Active badge and the published version at the top of the screen, and the agent's card in AI → Agents with the same badge.
After publishing#
- Test tab — ask it questions
- Design tab — adjust and publish again
- Activity tab — every run, with steps, cost, and what it proposed
- Schedule tab — make it run on its own
- Memory tab — see and delete what it remembers
- Pause / Activate — take it offline without deleting anything
- Duplicate (copy icon, on the card in the list) — the copy is always created as a draft
In the agent list, each card's Chat, Configure, and Activity buttons open the same screens directly, without going through the tabs.
Who can do what#
| Action | Who can |
|---|---|
| See the list | Anyone in the workspace |
| Chat | Member or above |
| Create, configure, and publish | Administrator or above |