Building agents
An agent is a conversation partner with its own instruction, its own documents and only the tools you select for it. It runs only when someone addresses it in the chat or writes to it in Microsoft Teams. If a task should start by itself or have fixed steps in a fixed order, build a workflow.
Create an agent
- In an empty chat, click Create agent.
- Describe the task in your own words.
- In the card Agent created, click Keep building.
The chat creates only the name and the description. The build chat on the agent's page then sets the instruction, tools and model. The button Create agent top right above the list Agents also leads to the empty chat.
Without a description, use + > New agent in the tile row below the input field of the empty chat, or … > Create empty agent top right above the list Agents. Both immediately create an empty agent and open its page.
Whichever way you choose, a new agent is private at first.
Find an agent
Agents in the sidebar opens the list of all agents and workflows. Agents carry the chip In the chat, workflows the chip With trigger.
To open an agent's page, click Edit. It is in the row menu of the list and in the chat header. Pin to sidebar in the row menu pins an agent to the sidebar.
Set up an agent in the build chat
An agent's page has a form on the left and a column with Build and Test on the right.
Under Build, describe in your own words what the agent should be able to do. The build chat then changes the name, description, icon, instruction, conversation starters, web search, tools, model and skills. While it works, the form is locked. Afterwards, the changed fields are highlighted briefly.
The build chat cannot create new skills and cannot assign MCP tools. You do both by hand.
The form saves every change automatically. There is no save button.
Test an agent
Under Test, you chat with the agent. This trial does not appear in any history. Reset clears it.
Set the model
In the row Model, you set the language model the agent answers with. It applies in the chat and in workflow runs in which a step uses this agent.
Every selectable model computes in the EU. The list is grouped into the tiers Fully sovereign and European servers. How the tiers differ is described under Sovereignty and models.
The top row is called Automatic. With it, the agent uses the default model of the workspace. If the workspace has not set one, a model of the tier Fully sovereign answers. A model of the tier European servers has to be chosen explicitly.
In the chat, anyone can override the model for the open conversation.
Add tools
Tools are individual actions of a service that the agent may call.
- In the card Tools, click Add.
- In the dialog Select tools, tick a whole service or individual actions.
If a service is not connected, its tools report an error. How to connect a service is described under Connecting services.
Approvals
Actions that change or send something carry the badge Writing. They need your approval before every call. Reading actions run directly.
The column Approval switches this per action. If you switch off the approval for a writing action, you confirm it with Yes, without asking.
In the chat, the switch Auto mode executes actions that require approval without asking. It appears as soon as the agent has tools, and it also applies under Test. The browser remembers it per agent. It confirms only calls to connected services. Cards of lowcloud itself, such as Add to knowledge or Start workflow, always wait for you.
What approving and rejecting trigger is described under Approvals and questions.
Web access
The group Web access in the card Tools has four switches:
- Web search: searches the web live.
- Web pages & files: fetches a URL you name, as a page or as a linked file.
- Read pages (Tavily) and Crawl website (Tavily): run at Tavily in the USA.
Web search and Web pages & files run on European servers and carry the sign 🇪🇺 EU hosting. The two Tavily switches carry 🌐 Worldwide.
Assign MCP tools
An MCP server is an external tool server that the workspace connects via a URL. It brings its own tools and its own access. How to connect a server and select its tools is described under Custom integrations and MCP.
On the agent, every connected server appears in the dialog Select tools as a service of its own. A tick on the server name switches on all its tools.
A tool marked Read-only on the server's detail page runs without approval. All others count as Writing and ask first. You switch off the approval as for any other writing action.
lowcloud itself is not an MCP server. Other assistants cannot use your agents as a tool.
Attach knowledge
In the card Knowledge, you attach documents the agent relies on when answering. The switch Use knowledge decides whether the agent also gets the global knowledge of the workspace. How both work is described under Knowledge.
An agent has no memory across conversations. Anything that should apply permanently belongs in the knowledge or in a skill.
Attach skills
A skill is a named instruction that the agent fetches when its description fits the task. You attach skills in the card Skills. How to create a skill is described under Skills.
Subagents
A subagent is another agent that this agent may commission in the chat. Several commissions run in parallel. The answer comes back as a tool result.
A subagent only reads and changes nothing. It cannot ask questions back and has no subagents itself.
Add conversation starters
Conversation starters are clickable suggestions in the empty chat of this agent. Add starter creates a new one.
Share and delete an agent
The row menu of the list Agents contains Share as link, Export and Delete. You can delete an agent only there. The agent's page has no delete button.
How to make an agent visible to the workspace, restrict access or pass it on as a copy is described under Sharing and access.
Limits
| Limit | Value |
|---|---|
| Tools per agent | 300 |
| Subagents per agent | 8 |
| Conversation starters per agent | 8 |
| Characters per conversation starter | 255 |
Permissions
Anyone who may build (the role admin, or the former role developer) can create and change agents. People with the role member chat with agents from the workspace's catalog. They do not see Edit. If they open an agent's page, they land on the list Agents.
In a partner's client workspace, only the partner's team builds. The client's people see no build functions there, not even with the role admin. Build pages redirect them to the list Agents, see Partner.