Modules
Agents
Base model, instructions, and parameters, published to the workspaces and people you choose.
On this page
An agent is a governed package of model, instructions, and parameters, published only to the workspaces and people you choose. It standardizes the model's behavior in chat and in flows, in a reviewable and auditable way.
Anatomy of an agent
| Field | Description |
|---|---|
| Name and description | How the agent appears in chat. |
| Base model | A local model or one from an external provider, enabled in every workspace the agent is published to. |
| Instructions | The system prompt: role, tone, answer format, and limits. |
| Generation parameters | Temperature, maximum tokens, repetition penalty, top-p, and top-k. |
| Status | Draft, published, or archived. |
| Access | Workspaces and people who can use the agent. |
Knowledge and actions
An agent carries neither its own knowledge context nor tools. Knowledge comes in two ways:
- In chat: from the knowledge or data context the person selects in the conversation.
- In a flow: from blocks such as Storage search and Data query, placed before the agent block.
For an agent to act on other systems, place it in a flow with Integration or Data write blocks.
Planned: tool use by agents and connection to MCP (Model Context Protocol) servers. These capabilities are not part of version 3.7.
Create and publish
- In Agents, create the agent and choose the base model.
- Write the instructions. Be explicit about what the agent should and should not do.
- Adjust the generation parameters, if needed.
- Select the workspaces and people with access.
- Publish. Draft or archived agents do not appear in chat.
Agents are created and published by administrators.
Use in chat and flows
A published agent appears as a conversation target in Chat for those who have access to it. In flows, the agent block sends the previous step's context to the agent and passes the answer to the following steps.
Every model call made by an agent goes into the unified log, with model, workspace, person, tokens, and rating.
Good practices
- One agent per task: focused agents are easier to review and audit.
- Explicit instructions on format, limits, and what to do when information is missing.
- For answers grounded in documents, combine the Storage search and Agent blocks in a flow.
- Prefer local models when the content must not leave your organization's environment.