Background Agents Need Endings

The new AI work surface is starting to look less like a window and more like an unfinished thought.

It follows you.

OpenAI has brought Codex into the ChatGPT mobile app so people can monitor active work, approve commands, review diffs, redirect threads, and move between connected machines from a phone. The latest Codex release notes add goal mode, appshots, browser annotations, and locked computer use, all pointed at the same direction: longer-running tasks that keep moving while the human is somewhere else.

Google is making a similar argument through Antigravity 2.0, with parallel agents, scheduled tasks, and background automation. Gemini Spark is framed as a helper that can work around the clock. GitHub Copilot sessions can now be monitored and steered remotely, with persistent debug logs and notifications from background commands. Copilot cloud agent can research, plan, and code on a branch before the work becomes a pull request.

The pattern is clear enough.

AI assistance is becoming less session-bound.

For years, most software respected the shape of a task because the task lived inside an active interface. You opened the document, edited the spreadsheet, reviewed the pull request, ran the command, closed the tab. The boundary was imperfect, but at least the object in front of you gave the work a kind of physical address.

Background agents loosen that address.

The work can begin at lunch, continue on a locked machine, ask for permission during a commute, send evidence to a phone, wait for a decision, resume on a remote host, and eventually arrive as a diff, a branch, a plan, a report, or a subtle obligation to review what happened while you were not fully there.

This is useful.

It is also a new attention contract.

The hype version of background work focuses on continuity. Nothing stalls. The idea you had while walking can become a task before it fades. A bug investigation can keep moving while you are in a meeting. A model can reach a fork, ask for direction, and avoid wasting an afternoon on the wrong path. The human gets to dip in at the right moments instead of babysitting the machinery.

There is real value in that.

But continuity has a shadow. If the tool can always keep going, the person may always remain partially available to it.

That is not the same as productivity.

It is a redesign of interruption.

The older interruption was a notification from another person or system: a message, an alert, a meeting reminder, a failed build. The new interruption is stranger because it may come from work you intentionally delegated. The agent is not distracting you from the task. It is the task, returning in fragments. A decision point. A permission request. A test failure. A screenshot. A branch that needs judgment before it can become real.

This makes the emotional texture different.

It can feel responsible to answer. After all, the work is already in motion. The agent is waiting on you. The task might finish if you just glance at the phone, approve the command, choose the approach, or skim the diff. The interruption arrives dressed as progress.

That is exactly why it needs design.

Good background agents will need better beginnings, certainly. Clear goals. Good source context. Explicit success criteria. Permission boundaries. Evidence trails. But the more neglected problem may be endings.

When is the task done enough to stop thinking about it?

When should an agent wait quietly instead of escalating?

When should a thread expire?

When should the system say: this needs full attention, not a mobile approval between errands?

When should the humane choice be to pause the work until the human is actually back inside the work?

These are interface questions, but they are also health questions. Modern work already has a boundary problem. The calendar leaks into the evening. The inbox leaks into the weekend. The unread thread becomes a low-grade pressure system. A lot of knowledge workers have learned to live inside a constant fog of small unfinished commitments.

Background agents could make that fog thinner by handling real work.

They could also make it denser by creating more partially active commitments than a person can metabolize.

The difference will not come from model capability alone.

It will come from the design of closure.

A mature agent interface should make pending work feel accountable, not ambient. It should show what is blocked, what is merely running, what requires judgment, what can safely wait, and what should be archived because it no longer matters. It should distinguish a reversible approval from a consequential decision. It should make it easy to stop a task without feeling like you have abandoned something living. It should treat silence as a design state, not a failure to maximize engagement.

This is where the word "assistant" becomes misleading. Assistants have social gravity. They ask. They wait. They make progress. They create the subtle sense that someone is depending on you, even when that someone is a system executing in a sandbox.

If every delegated task can return as a tiny claim on attention, the future of work may not feel automated. It may feel haunted by work in progress.

That sounds dramatic until you look at how much of work already consists of unfinished loops. The unreviewed draft. The unresolved comment. The almost-ready plan. The ticket waiting for context. The agentic version does not eliminate those loops. It accelerates their production and moves them into more places.

The question is not whether background agents should exist. They will, and in many cases they should. Some work genuinely benefits from time, persistence, and intermittent human steering. The question is whether the systems around them can preserve human attention as something more valuable than a routing layer.

The best tools will not only help us start more work from anywhere.

They will help us know what deserves to continue, what deserves review, and what deserves a clean ending.

Because a life filled with active threads is not automatically a life with more agency.

Sometimes agency is the ability to close the loop.