How to set AI agent autonomy levels for startups

AI agent autonomy levels are the per-capability switches that decide whether an agent must ask before it acts, or may run and leave a record. In Nautis those switches live under Chief of Staff for Owners and Admins: Ask every time, or Run without asking (auto-run). Everything defaults to asking. Deleting stays locked to always ask.

This is product education from Chief of Staff and AI Governance help as of October 2026. It is not permission to turn every External capability to auto-run on day one.

What AI agent autonomy levels control

AI agent autonomy levels apply the same rule everywhere agents act: chat, voice, briefing "Do it", routines, background agents, and connected AI apps. Ask every time means you see a card in chat or voice, or an Approvals entry for background work, and nothing runs until someone decides. Run without asking means the agent executes, the event lands in AI Activity, and you can undo within ten minutes where undo is supported.

Both paths are audit-logged. Chat lookups that only read data do not create activity rows, because they change nothing. If you cannot name who clears Approvals the same day, you are not ready for auto-run on anything that leaves the company.

External versus internal is the first cut

Treat the External column as a hard gate for the first weeks. Internal auto-run can wait until you have watched a week of Approvals with almost no rejects caused by wrong recipients or wrong boards.

A starter policy by job, not by vibes

Walk the Autonomy list once with a cofounder and write the policy in plain language before you click Save.

1. Drafting internal summaries and ranking briefs: candidates for auto-run only after the brief has been trustworthy for several mornings
2. Creating tasks or CRM rows inside Nautis: ask until stage names and boards are stable
3. Sending email, chat messages, or social posts: ask every time unless you have a narrow template and you review AI Activity daily
4. Anything that moves money language, pricing, or legal text: ask every time
5. Connected AI apps writing into the workspace: same rule as native agents; those writes are logged too

Per-action overrides exist when one action inside a capability needs a stricter rule than its siblings. Use Filter actions to find it instead of loosening the whole capability. New actions follow the new setting; past entries stay in the log.

How AI Governance proves what actually ran

Open Chief of Staff → Settings → AI Activity (Owners and Admins only), or the AI Governance page with shortcuts into Approvals and Autonomy. Anyone else who reaches the log sees that only owners and admins can view it. Newest rows first. The list loads twenty-five entries at a time; use Load more for older ones.

Each row shows the action in readable words, status, which agent acted, which model was recorded when one exists, and when it happened. Status filters cover All, Auto-ran, Approved, Rejected, and Pending. Under All you will also see Expired, Cancelled, or Logged when those apply. Click a row for sources, outcome, output, prompt, and approval or agent-job reference IDs when those were stored. Secret values from Credentials Vault tools are stripped from what gets logged. The screen is read-only: review here, approve in Approvals.

A useful Friday habit is filtering Auto-ran and reading every External row. If you cannot explain one of them, set that capability back to Ask every time before Monday. An empty Model column is normal when the action did not call a model or none was recorded.

When auto-run is rational

Auto-run earns its place when three conditions hold at once: the action is reversible or low blast radius, the inputs are stable (same boards, same templates), and a human still reads the activity log on a fixed cadence. Morning brief ranking that only reads existing workspace data is a different risk class from an agent emailing a customer.

Do not confuse speed with maturity. Approvals that feel slow usually mean too many routines proposing outward writes, not that Ask every time is broken. Pausing noisy routines fixes the queue faster than flipping External to auto-run.

Undo windows and silent failure modes

Where undo is supported, ten minutes is not a strategy; it is a last seatbelt. If nobody watches AI Activity, the window closes on mistakes nobody noticed. Auto-ran External actions that look "fine" in the log can still be wrong for the business (right text, wrong audience). That is why Show details on Approvals cards matters while you are still on Ask every time: you rehearse reading the blast radius before you remove the gate.

Logged status without Approved or Auto-ran still deserves a click. Open the entry; an Approval reference means it went through the approval process even if the filter label looks unfamiliar.

One afternoon that resets the dials

Leah left External send on auto-run after a clean week of internal task creation. A routine posted a pricing note to the wrong customer segment because a filter in the source board had drifted. AI Activity showed Auto-ran with the agent, timestamp, and output. Nothing sat in Approvals because nothing waited. She set every External capability back to Ask every time, paused the routine, fixed the board filter, and required Show details review for a full week before any auto-run discussion returned. The lesson was not "never automate." It was "auto-run without a same-day log review is how silent mistakes ship."

Connected apps inherit the same dials

Writes from connected AI apps such as Claude or ChatGPT are logged in AI Activity when they change workspace data. Do not assume a tighter Autonomy setting inside Nautis is irrelevant because the prompt started in another product. Set External capabilities to Ask every time while you are still learning which connected workflows create rows. Read Connect Claude, ChatGPT and other AI tools alongside Autonomy if those integrations are on.

Members who are not Owners or Admins can use many AI features without ever seeing the activity log. That split is intentional: operators propose, admins set policy and review. If only one admin exists, that person owns Friday Auto-ran review; do not leave the log orphaned during travel.

Translating policy into the Autonomy screen

Open Chief of Staff → Settings → Agent Autonomy. For each capability, read the External or Internal badge before you touch the dropdown. Prefer changing one capability, saving, then watching Approvals and AI Activity for a day before the next change. Bulk-loosening feels efficient and produces the pricing-note failure mode above.

When a capability offers Per-action overrides, tighten the dangerous action rather than the whole group. Example pattern: allow internal note drafting to auto-run while forcing ask on any action that attaches outbound recipients. If the UI copy says deleting always asks, believe it; do not hunt for a workaround.

After you save, send one low-risk test through chat that would create an Approvals card. Confirm the card text matches what you expect before you schedule a routine that depends on the same capability.

Autonomy next to Approvals, Routines, and briefing priorities

Autonomy without a cleared Approvals queue just moves risk into the backlog. Owners and Admins should open Approvals daily while any External capability is not on Ask every time. Each Approvals card shows which agent asked, when, and what would happen; cancel withdraws a request; unattended requests expire.

Routines that propose writes still obey the same Ask versus auto-run rule as chat and voice. Try a routine once before you schedule it. Test runs still ask before writes when the capability is on Ask every time, and they still use credits.

Briefing generation itself uses a very small amount of AI credits for the headline and still produces items if credits run out, with a plain fallback headline. Voice, chat, and routine writes use their own credits. Autonomy policy is about authorization, not about whether the brief appears; do not loosen External send because you are trying to "save" Approvals time when the real fix is fewer proposing routines.

Below Autonomy, What matters most right now holds six briefing priority sliders: Revenue at risk, Runway and burn, Deals and fundraising, Team health, Marketing and brand, and Everything else. Those weights shape tomorrow's briefing and also learn from done, snooze, and dismiss behavior. They are not a substitute for autonomy policy. Save them when focus changes; they apply from the next morning onward.

Chief of Staff is where Approvals and Autonomy live. AI Governance is the activity log that shows what ran, how it was authorized, and which agent did it. Routines are a common source of background proposals, so new routines should stay on ask until their transcripts look boring.

AI agent autonomy levels are a control surface, not a personality setting. Start strict on External, prove the log habit, then loosen one reversible internal job at a time. If a capability ever surprises you in AI Activity, the correct move is Ask every time first, debate later.