Otto asks at most once per task, and only when two readings would genuinely lead to different work. For larger side-tasks, it delegates to a worker that has read-only tools only.
Contents
Docs overviewThe conversation · Follow-ups & delegated workers
Asking, not guessing
Two built-in mechanisms keep the conversation honest: the follow-up question, when a decision could otherwise go too far wrong, and the delegated worker, when a side-task needs its own, thorough investigation.
When Otto asks
-
Only at a real fork
Otto only asks when two readings of the task would lead to genuinely different work — and the wrong one could not be fixed in two minutes. Anything it can look up itself (memory, contacts, mail, files, calendar) it resolves without asking.
-
At most one per task
A second follow-up within the same task is rejected by the tool — Otto then proceeds with its best assumption.
-
A menu or free text
Usually two to four clickable answers, up to twelve for multi-select questions; free text is always possible alongside.
-
A two-hour wait limit
If you do not answer, the question expires after two hours — Otto then proceeds with a stated assumption, never with an invented answer.
Every follow-up remains visible in history afterwards as a card — showing what was asked, and what you chose, typed, or let expire.
The delegated worker
- Its own reasoning run
- For demanding side-tasks, Otto delegates to a worker with its own brief and its own tools — Otto itself stays on a fast model for the fluid conversation.
- Read-only tools only
- A worker gets read-only tools only — it never sends, deletes, pays or acts on its own. Every one of its tool calls still passes through the same approval gate as any other task.
- At most four at once
- More workers running at once would flood the conversation with results — four cards are still readable, eight look like a wall.
- Truncated, never silently
- A very long result is truncated to 20,000 characters for Otto — visibly marked, pointing to the full version at the worker itself. For large tables, a separate, purely factual short summary is generated alongside, with no interpretation added.
- Chainable
- One worker’s result can be passed on unchanged as the next worker’s brief — useful where Otto would otherwise have to retype or paraphrase it itself.
Known gaps in this documentation build (as of 25 Sept 2026)
- Renaming a conversation is built in, but not yet covered by its own automated test.
- Asking a follow-up question to a worker that is still running is likewise untested — only cancelling a running worker is checked.
- The stop button that immediately halts a running task is explicitly listed as untested in the same inventory — and flagged as the riskiest of these three points.