Two Types of Agent Authorization Methods

LangSmith Fleet offers two authorization methods depending on whose authority an agent authenticates as when calling a tool.

The first is Assistant, an 'On-behalf-of' method that acts on behalf of the user. It maps permissions based on the ID of the user interacting with the agent. For example, it maintains security by restricting the agent's access to only the Notion pages or Slack data that the user themselves can access.

The second is Claw, a method in which the agent has its own fixed credentials. The agent's creator creates a dedicated account for the agent to control its permissions. This ensures that even when the agent is shared with external users, the creator's personal information is not exposed, and all users end up using the same permissions assigned to the agent.

Each method differs in the Channels it supports and its operational characteristics.

• Assistant: Since it requires mapping a user's channel ID (e.g., Slack ID) to their LangSmith ID, it is currently supported only on specific channels.
• Claw: Since it uses fixed credentials, it can be exposed across various channels. However, when performing sensitive tasks, it is recommended to verify execution through safeguards such as Human-in-the-loop.