Claude Lands in Google Docs With Approval Cards, ChatGPT Plans Ads for the Image-Generation Wait, and Gemini Agents Keep Working After You Close the Laptop

Three stories this week are about the same design question: when an assistant acts, waits or keeps working, what does the user see and control? We could not open primary pages from our environment, so details come from search summaries and secondary coverage, and we flag where sources disagree.

1. Claude Enters Google Docs, Sheets and Slides With Approval Cards

Anthropic's Claude for Google Workspace add-on and connectors entered public beta on October 6. Per Anthropic's announcement and secondary coverage, the default mode is "Ask before edits": each change appears as an approval card with a summary, and Claude waits. A second mode, "Accept all edits," lets it finish a task without pausing. In Docs, larger rewrites arrive as suggestion cards that can be applied or dismissed. One Taiwanese outlet and a Japanese blog report that links, external images and sharing or access changes still need separate approval in either mode; we could not confirm that against Anthropic's own text. The add-on is open to paid plans, and Team and Enterprise owners must enable the connectors first.

The lesson for builders is that safe defaults are a product feature. If your chatbot writes to a user's data, make the default a reviewable proposal and make the unattended mode an explicit opt-in. A minimal shape:

type EditMode = 'ask' | 'auto';

async function applyEdit(edit: ProposedEdit, mode: EditMode) {
  const risky = edit.kind === 'share' || edit.kind === 'external-link';
  if (mode === 'ask' || risky) {
    return queueForApproval(edit); // renders an approval card
  }
  return commit(edit);
}

Note that risky actions bypass the mode switch. Per-action risk classes age better than one global toggle.

2. Ads Move Into the Image-Generation Wait

OpenAI announced on October 5 that it will test visual ads that appear while ChatGPT generates an image. Reports put the US test in late October with an initial group of advertisers. OpenAI says the ads will be clearly labeled, separate from the image being created, and will not influence answers. Sources disagree on plan eligibility: one report says Free and Go users, others say OpenAI has not named the plans, and none we found explain frequency or whether users can hide the format.

Image generation is a natural slot because it has a built-in waiting period. For anyone building a monetized assistant, the design constraint is separation. Keep ad content out of the model's context, render it in a distinct, labeled component, and log impressions independently of the conversation so the answer path never depends on the ad path.

3. Persistent Gemini Agents Make Checkpoints a Requirement

At its Gemini at Work event on October 8, Google Cloud previewed a universal Gemini agent for enterprise work. Coverage describes a cloud-persistent agent with shared memory across devices, work that continues after a laptop closes, and temporary sub-agents for parallel steps. It is in private preview with no wider date, and industry versions for financial services and legal are also in preview. One outlet notes that long-term memory and multi-day execution already existed on Google's enterprise platform, so the news is mostly about reaching everyday work tools.

Once a task runs for hours, a dropped connection stops being a failure mode and a resumed session becomes the normal case. Persist progress as you go, and let the user check in without restarting the job:

await store.save(taskId, { step, state, pendingApprovals });
// On reconnect: resume from store.load(taskId), surface pendingApprovals first.

Conclusion

The thread connecting all three is control surface design: approvals for edits, labeling for ads, resumable state for long jobs. Models get the headlines, but these decisions shape whether users trust what you ship. Start with reviewable defaults, keep monetization out of the answer path, and treat every long task as one that will be interrupted.

A diagram of the approval-card flow (proposal, review, commit) would suit this post well.

— Maya