Skip to content
Atoi
Open Atoi
On this page

Asks, approvals, and safety

Understand when Atoi requests judgment, when it requests authority, and which actions keep a permanent human gate.

availableVerified Aug 24, 2026all

Atoi should act autonomously on routine, reversible work and pause when missing judgment or authority would make the result unsafe or meaningfully different.

#Ask chooses direction

An ask requests information or judgment. It should name the question, provide the relevant context, and explain the effect of each available choice. Your answer becomes part of the conversation.

Examples include choosing the intended audience, resolving two viable design directions, or deciding which source of truth should win.

#Approval authorizes an action

An approval presents a known consequential action before it runs. It should identify the exact object, destination, or scope. Approval is not a vague request to “let the agent work.”

Examples include publishing a public link, sending a final external message, changing permissions, or deleting meaningful data.

#Permanent gates

Some boundaries should not disappear behind a blanket autonomy setting:

  • an external send, post, upload, or submission shows the final payload;
  • deletion, permission, and settings changes identify the exact target;
  • credential, payment, passkey, and multi-factor steps require human control;
  • destructive or irreversible operations do not inherit a broad standing grant.

Atoi can remember narrower preferences where the product supports them, but a remembered rule does not silently widen these boundaries.

#Read the state literally

An unanswered ask means work is waiting, not failed. A denied approval is an explicit outcome, not permission to retry through another surface. An expired or superseded request should say so instead of accepting a stale decision.

The same logical request may appear in the conversation, Inbox, a native client, or a supported channel. Resolving it once should update every surface through the shared record.

#Review the receipt

After an allowed action, the conversation or task should show what ran, why it was allowed, and what result returned. For durable execution, continue with Tasks and receipts.

Verification sources
  • docs/decisions/2026-08-22-ask-user-interviews.md
  • docs/decisions/2026-08-22-permissions-gate-model.md
  • docs/decisions/2026-07-18-governed-work-contract.md

Was this useful?