Approvals and sensitive actions

Not every integration action should run unattended. Reading a Notion page is low-stakes; sending a Gmail reply on your behalf is not. Claos handles this with a two-state approval model: low-risk actions run automatically, sensitive actions pause for your sign-off.

What "sensitive" means

The integration catalog classifies each action by risk. The exact tiers vary by connector, but the dividing line is consistent:

Low-risk actions — read mail, list channels, fetch a page, search records. The agent calls them and continues.
Sensitive actions — send mail, post a message, create or delete a record, transition a ticket. These pause for approval by default.

You'll see this most often with messaging and ticketing connectors. A research agent that only reads from Notion will rarely hit an approval; a customer-reply agent that drafts and sends mail will hit one every run unless you change the policy.

Two surfaces, two flows

Approvals look different depending on where the agent run started.

In a chat

When you're chatting with an agent live and it tries a sensitive action, the chat pauses with a held action card showing what it wants to do — which integration, which action, the arguments it plans to send. You either Approve (it runs immediately and the agent picks up where it left off) or Decline (the agent gets back a "denied" signal and adapts).

This is the right experience for live work: you see what's about to happen, and the agent waits for you.

On a scheduled task

Background runs can't pause for a live user — no one's watching. So the behaviour inverts. The scheduled run ends when it hits a sensitive action, the held action gets persisted, and the schedule's status changes to Paused for approval with the integration label visible on the schedule card.

From there you can review the held action in chat the same way you'd review one from a live run, or you can change the schedule's approval policy.

The per-schedule policy

Open any scheduled task and you'll find an Approvals section with two settings:

Pause for approval (default). Sensitive actions stop the run and wait for you to review. Safe; sometimes noisy.
Auto-approve sensitive actions. The agent runs every action without pausing. Faster; only set this for schedules where you trust the agent's behaviour at this point.

A live preview shows you exactly which actions get auto-approved at the current policy — useful before you flip the switch.

The right default for a brand-new scheduled task is Pause. After a week or two of clean runs, flip to Auto-approve if the work isn't risky enough to justify the friction.

When to keep approvals on

A few cases where the friction is worth it even after weeks of clean runs:

Outbound communication. Sending mail, posting in customer-facing channels, posting in regulated channels. Even one bad send is more expensive than a hundred approvals.
Destructive actions. Deleting a record, archiving a ticket, closing a deal. Reversible in theory; expensive to reverse in practice.
Anything regulated. Legal, finance, healthcare integrations where the action has compliance implications. Keep a human in the loop.

Tips and gotchas

Approvals are per-schedule, not per-agent. Two different schedules using the same agent can have different approval policies — one paranoid, one liberal. The setting lives on the task's AI Schedule, not on the agent itself.
Live chat ignores the schedule policy. Held actions in chat always wait for your approval — there's no "auto-approve in chat" mode because you're already there.
A paused schedule doesn't keep retrying. The next scheduled tick runs whether or not the previous run is still waiting on approval. Old held actions persist independently until you act on them.
Disconnect on the third-party side too. Removing the integration from Claos stops new approval requests, but the OAuth grant in the third-party app still exists until you revoke it there.