How to Choose Which Agent Tools Need Human Approval in Arcade

November 21, 2026

Arcade lets you decide, per tool, whether the model runs it silently or asks the user first. That one setting is the difference between an agent people trust and one they mute after a week.

The default many teams ship with — approve everything — trains users to click through consent dialogs without reading. The other extreme, approve nothing, is how an agent sends 40 emails with the wrong signature.

Here is a policy that holds up.

What makes an action approval-worthy?

Three tests. If any is true, gate it.

Irreversibility. Sending an email, posting to Slack, creating a calendar event others can see, deleting a record, issuing a refund. Once it leaves, there is no undo.

Third-party visibility. Anything a person outside the conversation can see. Drafting an internal note is not the same as messaging a customer.

Money or identity. Payments, permission changes, anything that spends budget or grants access.

Reads almost never qualify. Listing Drive files, searching Jira, reading a calendar are safe to run silently as long as the tool only sees what that user can already see.

How do you keep approvals from becoming noise?

Batch related calls. If the agent needs to fetch five issues, that is one tool call the user should not see five times. Cache at the tool layer.

Show the payload, not the intent. A consent dialog saying “Run write tool?” is useless. Show the exact recipient, subject and body. Users approve what they can read.

Give a skip option. The third button should be “don’t do this” with the conversation continuing, not an error state that dead-ends the session.

Timeout sensibly. Ten minutes is long enough to read a draft; two hours means a queue of stale approvals the user will clear blindly.

In practice, a well-gated agent asks for approval two to four times per session, not twenty.

What do you log, and where?

Every gated call should record: user ID, tool name, arguments hash, approval timestamp, decision, and the provider response ID. Store the full arguments too — you will need them when someone asks “what did it send on Tuesday”.

Keep logs outside the model context. An agent that can read its own audit log is an agent that can reason about how to appear compliant.

Retention: 90 days is enough for most support workflows and long enough to cover a complaint cycle.

What breaks in week two?

Users start approving without reading. There is no fix other than making the dialog shorter — but the dialog is short because the action is small. If a call is too complex to summarise in one screen, it should be two calls.

Agents retry after approval. If a gated tool is called again in the same session with identical arguments, reuse the approval within a five-minute window. Otherwise people approve the same email twice and send it twice.

Multi-user threads. Two people in one thread means the wrong person may approve. Key approval to the acting user, not the conversation.

Who needs this policy layer?

Teams shipping agents to external users. Anyone in finance, healthcare or support where an email is a legal record. Startups that have already had one agent-action incident and are deciding whether to add guardrails or remove the feature — usually the guardrails are cheaper.

It is overkill for an internal single-user assistant and for demos, where a read-only tool set solves the problem by removing the risk rather than managing it.

Quick Summary

  • Gate any tool that is irreversible, visible to third parties, or spends money or grants access.
  • Let read-only tools run silently when they are scoped to that user’s existing permissions.
  • Show exact payloads in consent dialogs; batch related calls so users see 2–4 approvals, not 20.
  • Log user, tool, arguments hash and decision outside the model’s context, retained 90 days.
  • Reuse an approval for 5 minutes on identical retries to prevent double sends.
Tags: