Overview

An agent combines the instructions, model, product knowledge, and capabilities that power your embedded assistant. Its configuration determines how it responds, which information it can use, and which tasks it can perform on a user's behalf.

Create and configure agents under Agents in the Teaflask dashboard. Where the assistant appears and how it connects to your frontend are covered under Assistant UI.

Configure behavior

Each agent has its own settings for its role, responses, and audience:

Setting	What you configure
Prompt	The agent's role, responsibilities, tone, and instructions for responding to users.
Model	The model it uses, thinking effort, fallback models, and whether users may pick the model for their conversation.
Access	Whether it serves anonymous visitors or requires users to sign in to your app.

Choose capabilities

Capabilities give the agent access to product information and ways to act. You enable them on each agent, separately from writing its instructions.

Capability	What it lets the agent do
Knowledge	Search and read released documentation, with optional tag and folder restrictions.
Actions	Request selected API operations under the policies configured for those actions.
App access	Navigate your app, fill forms, and interact with the current page through your frontend integration.
Skills	Load reusable instructions relevant to a task.
Memory	Remember facts across conversations.
Delegation	Hand background work to subagents.
Workspace	Use a sandboxed computer.

Test and publish

Use the Playground to test your draft agent before making it available to users. You can adjust its instructions, model, and capabilities while the published revision continues to serve your app.

Publishing makes the selected agent revision available to users. An agent that has never been published is available for testing in the Playground, but does not serve your embedded assistant.

Revisions records changes to the agent's configuration and lets you compare versions. To roll back, restore an earlier revision and publish the restored configuration.

Agent configuration, documentation, and skills are published separately. Publishing an agent does not release draft docs or publish draft skills. Likewise, releasing documentation within an agent's existing knowledge scope does not require republishing the agent.

Continue reading

• Getting started: set up and embed your first agent.
• Knowledge: create and maintain the documentation agents use.
• Conversations: inspect messages, tool calls, and action results.