Skip to main content
Every action the agent takes is a tool card. Cards are the primary UI: they show what the agent did, what came back, and what state changed.

​
Card types

Each card maps to the same AgentSessionEvent sequence: tool_execution_start, one or more tool_execution_update, then tool_execution_end. The web adapter renders cards through the BEUI components in web/design.

​
Reading a card

  • Header — tool name, arguments preview, status (running, done, error).
  • Body — live output. For IPython this is stdout / stderr / display data. For shell it’s stdout / stderr. For edit it’s the diff.
  • Footer — timing, exit status, or a link into the subagent transcript for rlm cards.

​
Interactive questions

Some tool cards ask you a question instead of showing output. The turn pauses until you answer. In the web UI you’ll see an input; in the TUI you’ll get a prompt.

​
Building your own UI

If you’re rendering cards yourself, listen for the tool event on the stream and switch on the tool name. All card metadata is included in the event payload — you don’t need to import agent packages to render a card correctly.
Last modified on August 17, 2026