Ask something the model can't know on its own. The agent plans, calls get_current_time, shows the call and its result inline in the transcript, and folds the outcome into its reply.

Send one prompt to several models at once. Each reply lands as a branch on the same message — flip between them with the branch selector to compare answers side by side.

A run heading the wrong way doesn't have to finish. Interrupt it mid-stream, add a correction, and the agent picks up from there with the new instruction instead of starting over.

Every run leaves an activity trail. Opening a collapsed planning row reveals the full picture in the details panel — the exact prompt, the model's reasoning, and the action it chose.

Drill into each step of a run — title generation, decision planning, tool-parameter resolution — all the way down to the raw model response behind it.

Long history folds behind a checkpoint: earlier messages collapse into a summary that carries their context forward, keeping the working context lean without losing the thread.

A guardrail flags a risky tool call and pauses the run. Nothing executes until a human reviews the reason and approves — then the run continues exactly where it stopped.

Edit an earlier message and reprocess: the conversation branches at that point, and the selector switches between the original line and the new one — nothing is overwritten.

Attachments enter the context as compact summaries. When a follow-up needs detail the summary can't answer, an ask_attachment subagent reads the full file on demand.