Skip to main content
Unlisted page
This page is unlisted. Search engines will not index it, and only users having a direct link can access it.

Sessions

A session is a single, scoped conversation between an end user (or automated scheduled task) and the agents in your Chariot workflow. Sessions are the primary unit for tracking user interactions. They group all exchanges—messages, agent actions, and tool calls—under a unique context.

When a workflow is running, users can start new sessions to interact with the agents you have configured. Sessions can be initiated via the Chariot CLI or programmatically through integrated applications.

How Sessions Work​

  • Initiating a session:
    Users create sessions from an existing workflow. A new session represents a fresh conversation, subject to the workflow configuration.

  • Interaction flow:
    Messages sent by the end user (or scheduled task) are associated with the session. An agent handles the incoming requests and determines whether to respond directly, invoke tools, or delegate tasks to other agents.

Session Metadata​

Each session includes:

  • Session ID: Unique identifier
  • Timestamps: Creation and last activity times
  • Activities: Events recorded during a session (user messages, tool calls, etc.)
note

Sessions do not have names or descriptions.

Activity Types​

Sessions can include the following activity types:

User-initiated​

  • User to agent message: Text from an end user (or a scheduled task prompt) to a target agent.
  • Tool call approvals: Approval or rejection of pending tool calls.
  • Clear context: Resets an agent’s conversation context within the session.
  • Cancel: Stops in-flight work in the session.

Agent-generated​

  • Agent to user message: Text reply from an agent to the user.
  • Agent to agent message: Message passed from one agent to another; sometimes called a "handoff" in other frameworks.
  • Tool call requests: One or more tools the agent intends to run; this is generated right before a tool gets executed.
  • Tool call responses: Results after tools run (success or failure).
  • Thinking: Intermediate reasoning from the agent.
  • Structured output: Structured data from an agent (for example, artifact output).

Other​

  • Error: Details when processing fails.
  • Unknown: An event that does not match a recognized format.