Name the exact action class: draft, send, publish, purchase, browse, update, remember, or automate. Avoid vague approvals that say only "continue."
How to build consent receipts for personal AI agents
When a personal AI agent asks permission, the useful artifact is not only the answer. It is the evidence, scope, expiry rule, replay link, and rollback path that explain what the agent may safely do next.
Approval should leave a trail A receipt turns a fast yes into an inspectable operating boundary.
How-to workflow
Start with the moments where approval changes what the agent can do.
A consent receipt should be created when a personal AI agent crosses from conversation into consequence: sending a message, using private context, publishing a page, clicking through a browser workflow, spending money, or changing what it will do in the future.
Store the screenshot, draft, page state, price, recipient, diff, extracted fields, or summary the human reviewed before approving.
Record what is allowed, what is forbidden, which account or channel applies, and whether similar future actions are covered.
Add time limits for temporary context, sensitive recipients, travel, launch windows, customer issues, or volatile browser state.
Connect the receipt to a replay room or cached computer-use trace so later review does not depend on memory.
For public, financial, or external actions, store where the result went and how the operator can reverse or repair it.
Receipt schema
The fields that make approval reusable.
The receipt does not need to be heavy. It should be narrow enough for a machine to enforce and clear enough for a human to inspect later.
Use a compact structured record.
At minimum, store approver, timestamp, channel, requested action, evidence bundle, exact human response, permitted scope, excluded scope, expiry, result URL, replay link, and rollback owner. The transcript can remain attached, but the receipt should stand on its own.
For a text message AI assistant, this lets the text stay simple while the backend preserves the real permission boundary.
Evidence
What the user saw before saying yes.
Boundary
What the agent may do, may not do, and when it must ask again.
Recovery
Where the output landed and how to undo or repair it.
Browser agents
A computer-use cache can preserve the page state and screenshots that made an approval valid. If the page changes, the receipt should trigger reapproval.
Publishing agents
For AI agent website-building, store the page diff, planned live URL, publish status, source links, and rollback instruction.
Scrubbed principle
A consent receipt is the smallest durable object that can explain why an agent action was allowed, what evidence supported it, and when the permission should stop applying.
Approval surfaces
Route receipts by consequence, not by channel.
A good system can collect permission through text, browser UI, email, or a dashboard. The receipt format should stay consistent so the agent can enforce it no matter where approval happened.
Text
Fast approval, compact question, durable receipt.
Browser
Screenshot, state, selected control, changed page.
Publish
Diff, URL, source list, rollback owner.
Memory
Correction, learned rule, expiry, review date.
FAQ
Common implementation questions.
Consent receipts are not paperwork for their own sake. They are a compact safety primitive for agents that act outside the chat box.
Do receipts replace approval prompts?
No. The prompt is the front door. The receipt is the durable record created when the prompt is answered.
Should every user message create a receipt?
No. Create receipts for consequential actions: external communication, account changes, browser actions, publishing, private context use, spending, or learned permission changes.
How should receipts expire?
Use time, state, and scope triggers. Expire when the browser state changes, recipient changes, amount changes, public URL changes, or temporary context is no longer current.
Where does Super fit?
Super can act as the operator approval layer for personal AI agents, especially when approvals happen over text and need durable evidence.
Make every consequential yes inspectable.
A personal AI agent can move quickly without becoming opaque. Keep prompts lightweight, but turn important approvals into receipts the agent can enforce later.