Auto Mode

Let AI approve agent actions automatically so chats flow without click-by-click approval

Auto Mode lets a classifier auto-approve safe agent write actions so chats flow without stopping at every step. Each write is checked against what you asked for — actions that match your intent run through, anything outside it still pauses for approval. Reads always run free.

---

Why Auto Mode

Approval prompts protect you, but on a long chain of agent actions they slow chat to a crawl. Auto Mode keeps the safety net — a classifier reviews each write before it runs — and only interrupts you when judgment is actually needed.

• Anchored to your request. The classifier compares each proposed action against the message you sent. If the agent drifts off course, the action pauses for approval.
• Faster chats. Routine writes flow through; you stop click-clicking Approve.
• Still safe. Reads bypass the classifier entirely. Writes get reviewed one-by-one.
• Falls back to asking. When the action goes beyond what you asked for, or the classifier can’t decide, you see the normal Approve / Cancel prompt.
• Reversible anytime. Flip it off mid-conversation and per-tool rules resume immediately.

---

Turn it on

Two places — pick whichever is in front of you.

From a chat

Click the Auto mode chip in the composer.

The Auto mode chip sits next to the model selector in the composer.

From settings

Open Approval in workspace settings and use the Auto Mode toggle at the top.

Approval settings — Auto Mode banner sits above the permission tabs.

Auto Mode is workspace-scoped. Turning it on from a chat persists for every conversation in that workspace until you turn it off.

---

What you’ll see in chat

Auto-approved actions show a small “Auto-mode allowed this” line under the tool call. No button to click.

Each auto-approved action stays visible with a confirmation line — you can scroll back and see what ran.

Risky actions still show the regular Approve / Cancel prompt, just like with Auto Mode off.

---

How it decides

The classifier reads two things together: your original message and the specific action the agent wants to run. If the action is a faithful step toward what you asked for, it goes through. If it drifts — wrong resource, wider scope, or a verb you didn’t ask for — it pauses.

Action	What happens
Reads (list, describe, get)	Always run — classifier not invoked
Writes that match your intent	Auto-approved, “Auto-mode allowed this”
Writes that go beyond your intent (different resource, wider scope, destructive verb you didn’t ask for)	Still asks
High-risk writes (IAM, broad blast radius, irreversible)	Still asks even when intent matches
Same action that keeps failing	Blocks to prevent runaway loops

The classifier is the policy. You can’t edit its rules — that’s intentional so behavior stays consistent across workspaces.

---

Examples

Allowed	Asks anyway
You said: “please stop the staging EC2 instance” Agent ran: Stop EC2 instance on the staging instance Why allowed: The action matches the verb (stop), the resource type (EC2), and the specific instance you named. Stop is also reversible — you can start it again.	You said: “please stop the staging EC2 instance” Agent tried: Terminate EC2 instance on the staging instance Why it asks: Terminate is destructive and irreversible — you asked to stop, not delete. Auto Mode pauses and shows the normal Approve / Cancel prompt so you can correct course.

---

Auto Mode and per-tool settings

When Auto Mode is on, the per-tool Requires Approval flags on the Approval page are paused. Each write tool shows an orange ⚡ Classifier badge instead of Allowed / Requires Approval — that’s the classifier deciding case-by-case in place of the static rule.

Read-only tools keep their green Allowed badge and run as before.

Turn Auto Mode off and per-tool rules resume unchanged — nothing you configured there is lost.

Tools you’ve set to Disabled stay disabled even when Auto Mode is on. Auto Mode never enables a tool you’ve turned off.

---

Safety rails

Control	Description
Per-action review	The classifier looks at each write call individually, not the conversation as a whole
Audit trail	Every auto decision is logged so you can review what ran and why
Self-pausing	If the classifier itself becomes unresponsive, Auto Mode pauses and the banner switches to a paused state — you go back to manual approval until it recovers
Toggle anytime	Flip Auto Mode off mid-conversation; takes effect on the next tool call
Disabled tools stay disabled	Auto Mode never overrides a tool you’ve turned off

---

When to use it

Good fit	Leave it off
Dev and staging workspaces Exploratory chats where you’d otherwise click Approve dozens of times Long multi-step plans run by Anna Recurring tasks where you’ve already vetted the pattern	Production workspaces with sensitive resources First-time use of a new connection — build trust manually first Compliance-sensitive environments where every change needs a human sign-off

---

FAQ

Why did Auto Mode still ask me?

The classifier judged the action high-risk — usually destructive verbs, IAM changes, or actions with a wide blast radius. Approve / Cancel as normal.

Why did it block instead of asking?

Auto Mode blocks when the same action has failed several times in a row, to avoid runaway loops. Rephrase the request or fix the underlying issue, then try again.

Can I customize what Auto Mode auto-approves?

Not directly — the classifier rules are fixed so behavior stays consistent everywhere. If you want stricter control, leave Auto Mode off and use Tool Permissions on the Approval page to set per-tool rules manually.

Does Auto Mode override Disabled tools?

No. A tool you’ve set to Disabled stays disabled. Auto Mode only changes how approval decisions get made, never whether a tool can be used.

What if the classifier is slow or unavailable?

Auto Mode pauses itself and the banner switches to a paused state. Approvals fall back to manual until it recovers — no action runs without a decision being made.

Is Auto Mode per chat or per workspace?

Per workspace. Toggling it on from one chat persists across every conversation in that workspace until you turn it off.

---

Related

Approval

Per-tool permissions, approvers, and command permissions

Capabilities

What agents can produce — dashboards, reports, slides

Agents

How agents work and collaborate

Connections

Set up cloud and service connections