Background Agent
An AI agent that runs continuously on a remote server, executing tasks on a user's behalf even when the client device is closed.
What It Is
A background agent is an AI agent whose runtime does not live on your laptop or your phone. It lives on a server somewhere, attached to your identity, working on the instructions you gave it the last time you opened the app. When you close the lid, the agent does not stop. When you put the phone in your pocket and walk into a meeting, the agent does not pause. It keeps running, and it reports back the next time you check in.
The term has been around for years inside cron jobs and message queues, but its meaning shifted in 2026 with the consumer rollout of Google’s Gemini Spark, OpenAI’s Managed Agents, and a handful of similar products. In each of those products the background agent is not a script you wrote. It is a general-purpose reasoning loop that takes natural-language instructions, calls tools through Model Context Protocol connectors, and returns either a finished outcome or a question. The user does not have to be present. That is the new part.
How It Actually Works
Three pieces have to be in place for a background agent to be useful. The first is a persistent runtime: a server that holds the agent’s state, its tool permissions, and its current task between sessions. The second is a set of credentialed connectors to the apps where the work happens. Gemini Spark uses MCP to reach Canva, OpenTable, and Instacart at launch. Without those connectors, the agent has nothing to act on. The third is a notification channel back to the user, because an agent that cannot ask a clarifying question is an agent that will eventually make an expensive mistake.
Under the hood, the loop looks like read instruction, plan, call a tool, observe the result, decide whether to continue or escalate, repeat. The agent harness handles retries, timeouts, and the boundary between what the agent can do silently and what it has to surface for your approval.
Why It Matters Right Now
Until this year, the agent demos that mattered were happening inside developer tools: Claude Code, Codex CLI, Cursor’s parallel agents. Those tools assume the user is at a keyboard. The shift in 2026 is that the same architecture is being repackaged for a person who never opens a terminal. Spark, Universal Cart’s Agent Payments Protocol, and Klarna’s Shopping Search app in ChatGPT are all built on the same assumption: the user gives intent, the agent executes asynchronously, the user reviews the outcome. That is a different software pattern than the one most of the consumer internet was built around, and it is the pattern under most of the consumer AI launches of the last six weeks.
How TWO Uses It
The first question we ask when a tool ships with a background-agent mode is not what it can do. It is which of your decisions you will stop touching once you have it. A daily-digest tool that keeps fetching news while you sleep is useful. A purchase agent that ships an order to your house while you sleep is something else. The cost of the agent is not its monthly subscription. The cost is the gradual erosion of the small decisions that, in aggregate, make up your attention.
So at TWO we use a simple rule when evaluating a new background agent: name three tasks the agent will absorb, name one task you will deliberately keep doing by hand, and write both down before you turn the agent on. The second list is the one that protects you.
A Concrete Operator Scenario
You sign up for a $100-per-month managed-agent and you give it standing instructions: book Saturday dinner before noon Friday, refill the household groceries when the pantry list hits ten items, watch the airfare on a flight you might take and alert you if it drops below a number. Two weeks in, the agent has done all three competently. Your fourth instinct, to add it to the calendar-scheduling task, is the one to question. Calendar scheduling is where you would lose the ability to notice that two of next week’s meetings should be one phone call, which is a judgment call no ai-agent can make for you. The rule is to give the agent the repetitions and keep the judgments.