What human in the loop means for AI agents

Human in the loop AI agents, in Nautis, is the person who decides before an agent action runs when that capability uses Ask every time. The Chief of Staff help (checked 31 Oct 2026) defines Ask every time as: you see the action first, a card in chat or voice, or an entry in Approvals for background agents, and nothing happens until you approve. The same page names the Approvals view as the Approval Center. AI Governance (checked 31 Oct 2026) is the running record of what agents did, together with the Approvals tab and Agent Autonomy settings. This post is about clearing that human gate. How to set the Ask every time versus Run without asking dials by capability is a separate topic (setting AI agent autonomy levels); this post does not reteach that policy matrix.

Human in the loop AI agents means a Pending card, not a slogan

Chief of Staff FAQ: everything starts on Ask every time and lands in the Approval Center, with one exception: Quick capture (inbox) starts on Run without asking, because it only saves a note to your own private capture inbox. Owners and Admins can change any capability except AI Copilot deletes, which shows Always asks with a lock: some actions cannot be automated; deleting always asks (deleting records in Nautis). Anything set to Run without asking does not wait for approval and is logged in AI Activity.

Open the Approval Center and decide

1. Open the Approvals view, the Approval Center. Anything your AI agents want to do on your behalf waits here; the tab badge shows the pending count.
2. Each card shows which agent asked, when, what the action is, and a plain-language summary. Click Show details to inspect exactly what would be sent or changed.
3. Owners and Admins can Approve (confirm copy: your agent will execute this action now; this can have real effects), Reject, or Cancel request (anyone can cancel).
4. Switch from Pending to History and filter by All, Approved, Rejected, Expired, or Cancelled. Unattended requests expire on their own.

A card from your own chat or from Do it is yours to decide, whatever your role. Only organization Owners and Admins can approve or reject requests from background agents and change Agent Autonomy settings. Only Owners and Admins can view the AI Activity log.

Clear Pending the same day. If cards expire because nobody decided, the human loop failed even though Ask every time was set correctly.

Briefing and Do it feed the same Approval Center

The daily Founder Briefing groups items into Priorities, Blockers, Approvals, Pipeline, Today (meetings and calendar), and Runway. The At a glance panel shows pending approvals. The Dashboard Today's briefing card can show a needs approval chip.

Chief of Staff FAQ: Do it and approving in the Approval Center are the same pipeline. Tapping Do it does not approve anything, whatever your role. For a change, it creates a card in your Chief of Staff chat and the matching request in the Approval Center, and you approve it from there. The exception is an action set to Run without asking under Agent Autonomy, which runs right away. Drafts always get a card so you can review and edit them first. Lookups run right away and appear under the item.

AI Activity reviews what already happened

Open Chief of Staff, Settings, AI Activity (or the standalone AI Governance page, which adds shortcuts to Approval Center and Autonomy settings). Newest actions appear first. Each row shows Action, Status, Agent, Model where recorded, and When. Filter by All, Auto-ran, Approved, Rejected, or Pending. Entries can also show Expired, Cancelled, or Logged under All. Open a row for recorded When, Approval and Agent job references, Sources, Outcome, Output, and Prompt.

The log records actions, not every question. Writes, approval decisions, and background runs are logged. A chat lookup that only reads data does not add an entry. Writes from connected AI apps, such as Claude or ChatGPT, are logged too. Secret values stored through Credentials Vault tools are removed from what gets logged. The log is read-only with no edit, delete, or export button, and covers one organization at a time. Agent Autonomy settings state that every action, automatic or approved, is audit-logged.

AI Governance is explicit: where do you approve a pending action? In Chief of Staff, under Approvals. The activity log is for reviewing, so it has no approve button.

External, Internal, and Per-action overrides (pointer only)

AI Governance and Chief of Staff both state that each capability is labeled External (touches the outside world) or Internal (stays inside Nautis). Some capabilities also have Per-action overrides so you can treat one action differently from the rest; use Filter actions to find it. The same autonomy rule applies everywhere: chat, voice, briefing Do it, routines, background agents, and connected AI apps.

Choosing which capabilities stay on Ask every time versus Run without asking is the autonomy-levels job. This post assumes you already know those labels exist; keep the human loop healthy by deciding Pending cards and reading Auto-ran rows after the fact. For ask-first outbound email habits, see AI that asks before acting on email. Full screens: Chief of Staff and AI Governance.

An illustrative Pending clear (not a customer story)

The walkthrough below is an illustrative example only. It is not a real company log, measured result, or customer quote. Labels match the help pages above.

A founder opens Briefing and sees the Approvals section with pending items, plus a needs approval chip on the Dashboard card. In the Approval Center the badge shows two Pending cards: one Internal update from a routine, one External email draft from chat. They click Show details on both, Approve the update, and Reject the email. History shows Approved and Rejected. AI Activity later lists those statuses. Quick capture remains on Run without asking and only saves a private inbox note. AI Copilot deletes remains Always asks. No Pending card is left to hit Expired that day.

Pending, Expired, and History without mixing them up

In the Approval Center, Pending means still waiting for a decision. History lets you filter All, Approved, Rejected, Expired, or Cancelled. AI Governance uses the same Pending idea on the activity log: still waiting for a decision. Expired means nobody decided in time. Cancelled means the request was withdrawn. Approved and Rejected mean a person already decided. Auto-ran on AI Activity means the action used Run without asking and did not wait in Approvals.

Treat Expired as a process failure to investigate, not as a quiet no. If External sends keep expiring, either nobody is opening Approvals or the wrong people expect background cards to clear themselves. Owners and Admins own background approvals; everyone else still owns their own chat and Do it cards.

What the human actually inspects on a card

Show details exists so you can see exactly what would be sent or changed before Approve. Use it on External actions every time. On Internal edits, still read the summary: wrong lead stage or wrong note body is still a real effect after Approve. The Approve confirm copy states the agent will execute the action now and that this can have real effects. Reject declines. Cancel request withdraws; anyone can cancel.

Do not Approve from the briefing list alone. Do it only creates the card. The decision is on that card in chat or in the Approval Center.

Notifications and the morning chip

Chief of Staff includes a Notifications view for everything your Chief of Staff and routines did, plus delivery settings. The Dashboard Today's briefing card can show a needs approval chip beside runway and top items. Use that chip as a jump into the human loop, then finish in Approvals until the pending count is clear. The briefing itself is per-organization; done, snooze, and dismiss update it for everyone, but approving an agent action is still the Approvals or chat-card path described above.

Make the loop daily

When At a glance shows pending approvals, open Approvals before you chase pipeline. Decide, Reject, or Cancel. Skim AI Activity for Auto-ran if anything is on Run without asking. Human in the loop AI agents is that middle step between Ask every time and the audit line: a person in the Approval Center or on the chat card. Without it, you only have Expired and surprise Auto-ran rows.