An AI agent is an assistant you configure in AI Center: you choose its model, write its system prompt and decide which connectors it can use. To create one, go to AI Center → Agents, click Create agent, enter a name and the system prompt, choose a model, and click Create agent. You can then test it in the chat preview and share it with your users.
For example, a "Solar Copilot" can focus on inverter output and string faults, while an "Energy Analyst" reports consumption and efficiency across sites. Each one answers from the same Ubidots data, following its own instructions.
Requirements
A plan that includes AI Center. The number of agents you can create depends on your plan; see AI Center: Overview, Plans and AI Credits.
Account owner or admin role.
1. Create an agent
In the navigation menu, click AI Center → Agents. You'll see a card for each agent, with its model, label, name and description. Use Search agents... to find one.
Click Create agent. The New AI Agent page opens, with the Settings, Connectors and Advanced tabs on the left and a chat preview on the right.
On Settings, enter the Agent name, an optional Description and the System prompt. The system prompt is required.
On Advanced, choose the Model. You can also adjust Temperature, Max tokens and zero data retention there.
Click Create agent. You can then test the agent in the chat preview, attach connectors and keep adjusting it.
2. Agent settings
The name, description and system prompt are on the Settings tab. The model, temperature, max tokens and zero data retention are on the Advanced tab.
Setting | What it does | Limits and default |
Agent name | The name users see, e.g. "Solar Copilot". | Required. Up to 256 characters. |
Agent label | Generated from the name and used in API calls. | Read-only. It doesn't change if you rename the agent. |
Description | A short summary of what the agent does, shown on its card. | Up to 8,000 characters. |
Model | The AI model that powers the agent. Models are grouped as Premium, Balanced, Fast and Others, and each shows its context window and credit multiplier. | Required. |
System prompt | The agent's instructions: its role, what it knows about your operation and how it should answer. Sent with every message. | Required. Up to 8,000 characters. |
Temperature | How focused or varied the answers are. Lower values produce more focused replies. | 0 to 1. Default: 0.2. |
Max tokens | Upper bound for the length of each answer. If an answer reaches it, it's cut off with "Response truncated — the agent hit its max tokens limit". | 256 to 64,000. Default: 4,096. |
Chat icon and Primary color | How the agent looks in the chat. Reset to default restores them. | Optional. |
Starter messages | Suggested prompts shown to end users before they start chatting. Click Add starter message to add one. This tab appears when it's available on your account. | Up to 5, 200 characters each. |
Zero data retention (ZDR) | Whether the AI model provider may keep a copy of the data. See section 4. | On by default. Only the account owner can change it. |
Choosing a model
Each model has a credit multiplier, shown as Credits: N× per token. A model in the Fast group answers routine questions quickly and uses fewer AI credits; Premium models are better at long, multi-step analysis and use more. Balanced models sit in between and are a good starting point. See how AI credits work.
If you switch an existing agent to a different model, save it before testing it in the chat preview.
Writing a good system prompt
The system prompt is where the agent learns about your operation. A useful one usually covers:
Role: "You are the maintenance assistant for the cold-storage warehouses of ACME Foods."
Context: the sites, device types and variables that matter, their units, and the normal ranges or thresholds.
How to answer: language, length, format (for example, a short summary first and details after), and what to do when data is missing.
Limits: topics to stay away from, or when to tell the user to contact a person.
The system prompt is sent with every message, so keep it focused: a long prompt makes every answer use more AI credits.
3. Test the agent in the chat preview
The Chat preview on the right of the agent's page lets you talk to it before anyone else does. A new agent has to be saved first; after that, the preview runs in Testing mode — this conversation won't be saved.
Changes take effect when you save the agent, including in the chat preview.
Under each answer you see the tokens used, its cost, how long it took and which tools the agent called, so you can compare models and prompts.
Click Reset conversation to start over.
4. Privacy: zero data retention
On the agent's Advanced tab, the Zero data retention (ZDR) switch controls how the AI model provider treats your data:
On (default, shown as Active): prompts, device data and responses are processed to answer each request, then discarded. The AI model provider keeps no copy.
Off (shown as Off for testing): requests can also reach providers that keep data for a limited time. Use it only to compare models, and turn it back on before the agent goes live.
Whether ZDR is on or off, your data is never used by AI providers to train their models.
Note: Only the account owner can change zero data retention. If an admin changes it, the save fails and none of that save's changes are kept. If the agent's model doesn't support ZDR, the switch shows Not available; choose another model to turn it on. |
5. Connect the agent to data and tools
Every agent comes with the built-in Ubidots connector, so it can look up your devices, variables, values, statistics, events and incidents without any setup.
To give the agent data or tools from other systems, open its Connectors tab. It shows the connectors already Attached and those Available to attach; click Attach or Detach, then save the agent. You need to save a new agent once before you can attach connectors. This tab appears when Connectors are available on your account. See Connectors.
6. Share the agent
You share an agent through a channel: the web console inside your Ubidots app, or WhatsApp. Channels control access to agents: each channel's access rules decide which users get which agents. See Web Console for End Users and Talk to Your Industrial AI Agents through WhatsApp.
7. Edit or delete an agent
Edit
Click the agent's card, or open its ⋮ menu and click Edit. The Edit AI Agent page opens with the current settings, on the same tabs. Make your changes and click Save changes. If you leave with unsaved changes, you're asked whether to discard them.
Delete
Open the agent's ⋮ menu, click Delete and confirm. Deleting an agent can't be undone. Conversations already open with it stop accepting new messages but stay readable, and new chats fall back to the Ubidots agent. While it's being removed, its card shows Deleting….
8. Use an agent in Chat
Go to AI Center → Chat, click the agent's name below the message box and pick your agent under My Agents. Then click New chat and ask your question. See AI Center: Overview, Plans and AI Credits for tips on asking good questions.
Frequently asked questions
Which data can an agent see?
An agent works with the permissions of the user it's answering. In your own Chat, that's you; in a channel, it's the end user who is logged in, so each end user only gets answers about the devices and data they can already access.
How many agents can I create?
It depends on your plan. See the plans table in AI Center: Overview, Plans and AI Credits.
Why did an answer stop halfway?
It reached the agent's Max tokens. Raise the limit, or ask the agent in the system prompt to keep answers shorter.