How to build approval replay rooms for agents

A replay room is the operator-facing memory layer around an approval. It stores what the agent asked, what evidence the human saw, what the human allowed, and what future runs should remember before asking again.

Explore Super Design text approvals
evidence bundlesconsent scopereplayable handoffssafe resumesevidence bundlesconsent scope

Start by separating approval from memory.

The approval itself can be a short tap, reply, or confirmation. The replay room is the durable context that lets the operator, the agent, and a future reviewer understand the decision without reconstructing the whole session from raw logs.

The minimum viable replay room

A useful replay room starts with six objects: the agent request, the proposed action, the evidence bundle, the human response, the consent boundary, and the outcome. Without those pieces, the system can still claim that a person approved something, but it cannot confidently explain what the person believed they were approving.

For a text-message AI assistant, the request might be short enough to fit in SMS. The replay room should preserve the broader context that did not fit in the message: screenshots, document excerpts, payment amount, recipient, deadline, and rollback instructions.

Evidence wall for a replay room

Request

Write the agent request in human language. Avoid only storing internal tool names or opaque traces.

Evidence

Attach the artifacts the operator relied on: screenshots, diffs, extracted text, source links, and risk notes.

Boundary

Record what the approval allows, what it does not allow, and when the permission should expire.

Implementation checklist

  • Create a stable approval id before asking the human.
  • Snapshot the evidence bundle at the moment of approval.
  • Store the exact human response, channel, and timestamp.
  • Summarize the consent boundary in plain language.
  • Link any browser cache, file diff, or generated output.
  • Write a resume instruction for the next agent run.

What not to store alone

A bare event log, isolated SMS reply, or database boolean is too thin for high-value delegation. It may tell you that approval happened, but it does not tell a future agent whether a similar action is safe, whether the user was approving one item or a category, or whether the evidence has since changed.

Design the room around the reviewer, not the database.

Replay rooms become useful when they are organized around the questions an operator actually asks after a delegation: what was proposed, what proof was shown, what changed, and what should happen next.

Approval prompt

Show the original prompt and a short operator-readable reason for the ask.

Evidence panel

Include screenshots, citations, page state, and diffs. A computer-use cache is especially valuable here.

Consent ledger

Make scope explicit: single action, repeated action, spending limit, account boundary, or time window.

Outcome trace

For generated artifacts such as pages from AI agent website building, show the final URL, publish state, and rollback path.

Make approvals replayable before they become expensive.

Good replay rooms reduce repeated questions, improve operator trust, and give agents safer context when they resume work.

Capture the evidence, bind it to consent, then teach future runs what the approval meant.

Before the ask

Generate a proposed action and gather proof before contacting the operator. Do not ask for approval while evidence is still being fetched. The replay room should already know the relevant account, object, amount, recipient, page, file, or output before the human is interrupted.

During the ask

Send the shortest safe prompt over the user’s chosen channel, then link that prompt to the room. In SMS and iMessage workflows, the message can stay compact while the replay room keeps the full decision context available behind the scenes.

After the response

Convert the human reply into structured consent. A reply of “yes” should become a machine-readable boundary, such as approve this specific purchase, approve this publish action, or approve this browser session only for the next ten minutes.

At resume time

When the agent resumes, it should load the replay room before deciding whether to ask again. If the new action is inside the original boundary and the evidence has not materially changed, the agent can proceed or present a lighter confirmation. If the evidence changed, it should ask again with a fresh room.

FAQ for operators building approval memory.

Approval replay rooms are not only compliance artifacts. They are product surfaces that help personal agents become more predictable and less annoying.

Is a replay room the same as an audit log?

No. An audit log is usually chronological and technical. A replay room is decision-centered and operator-readable. It can link to the audit log, but it should explain the approval in plain language.

How long should consent last?

Consent should last only as long as the operator would reasonably expect. Tie it to one action, a narrow task, a time window, or a budget. Avoid broad standing consent unless the product has strong review and revocation controls.

What evidence is worth preserving?

Preserve anything that would change the decision if it were different: price, recipient, screen state, calendar time, file diff, generated page, source citation, or policy note.

Where does Super fit?

Super focuses on practical personal agent workflows across messaging, browser use, and generated outputs. Replay rooms are a natural design layer for that kind of delegated work.

Sources and reference anchors