Background Agents Need a Stop Button
2026-05-21 · 4 min read · Janaina Maia
The most interesting part of Google’s latest Gemini announcement is not that the assistant can answer more questions. It is that Google wants Gemini to keep working when you are not there.
Google introduced Gemini Spark, a personal AI agent designed to run in the background, connect to tools like Gmail, Docs, Slides, and future MCP connections, and complete tasks under the user’s direction. Google also announced Daily Brief, which gathers updates from connected apps and turns them into a prioritised morning summary.
This is a useful direction. It is also where AI product design gets much harder.
Background work changes the trust contract.
When an AI assistant answers a question in front of me, I can inspect the answer immediately. I can challenge it, ignore it, copy it, or ask again. The interaction is contained.
A background agent is different. It can wait, monitor, organise, draft, trigger, and continue while I am doing something else. That shifts the product from conversation to delegation.
Delegation needs a stronger contract than chat. If the agent is going to keep working after the laptop closes, users need to understand what it is allowed to do, when it will act, which tools it can touch, what counts as a risky action, and how they can pause or reverse the work.
The stop button is not a small feature.
Google says Spark is designed to ask first before high-stakes actions like spending money or sending emails. That is the right instinct. But the deeper design question is not only “does the system ask before the final action?” It is “does the user stay oriented while the work is unfolding?”
For background agents, I would want to see clear product surfaces for:
- Standing instructions: what the agent should do repeatedly, and what it should never do.
- Scope: which apps, folders, messages, and data sources are in bounds for this task.
- Checkpoints: where the agent must pause for human review before continuing.
- Activity trails: what it looked at, what it changed, what it drafted, and what it ignored.
- Stop and rollback: a simple way to pause the agent, cancel a run, undo a change, or narrow permissions.
These are not governance decorations. They are the interaction model.
Proactivity can easily become anxiety.
There is a fine line between a helpful assistant and a system that feels like it is moving around your digital life without enough consent.
Daily Brief is a good example of the opportunity. A well-designed morning brief could reduce cognitive load by surfacing what matters across email, calendar, and follow-ups. But if the brief is opaque, overconfident, or too noisy, it becomes another feed to manage. The value is not in summarising everything. The value is in making the user feel more prepared, not more watched.
That is the product challenge for proactive AI. The system has to earn the right to interrupt, prioritise, and act.
The enterprise version will be stricter.
In consumer products, a background agent might manage subscriptions, school emails, travel plans, or personal documents. In enterprise products, the same pattern touches customer records, engineering decisions, compliance evidence, financial workflows, and operational approvals.
The stakes are different. A background agent in an enterprise setting cannot only be charming and useful. It needs policy-aware permissions, audit trails, role boundaries, escalation paths, and visible accountability.
Design leaders should not wait for engineering to “add governance later.” The permission model is part of the experience. The review surface is part of the experience. The stop button is part of the experience.
My take.
The next wave of AI products will not be judged only by how well they answer. They will be judged by how safely they act when users are not staring at them.
Background agents are powerful because they promise to remove coordination work: monitoring, sorting, drafting, following up, and preparing. But the more invisible the work becomes, the more visible the controls need to be.
If an AI agent keeps working after I leave, I need more than confidence that it is smart. I need a clear agreement, a visible trail, and a stop button I can trust.