Quick Answer: Background agents are AI agents that keep working after you leave. You hand over a goal once; the agent runs on its own machine, wakes up on a schedule or an event, remembers the last run, and reports back with finished work.
Most AI still stops the moment you stop typing. Background agents don't. You give one a job — the Monday sales report, a weekly competitor price check, chasing overdue replies — and it keeps going after you close the laptop, then shows up with the result.
The idea started with coding tools. In September 2026 it moved into ordinary office work, when OpenAI, Microsoft, Perplexity and Salesforce all shipped agents built to run unattended. It's also the idea cowork.ink is built on: AI coworkers with their own computers, working on a schedule. Here's what background agents are, what each vendor shipped, and what an agent actually needs to work while nobody is watching.
What are background agents?
A background agent is an AI agent that works on a goal outside the conversation that started it. You don't wait for an answer. The agent takes the task, runs on its own machine for minutes, hours or days, and comes back with a result.
One note on terms: some chatbot platforms use "background agent" for a silent helper that watches a live chat and tags or routes it. This article means the other, more common sense — an agent you delegate a task to and walk away from.
| Foreground agent | Background agent | |
|---|---|---|
| Where it runs | Your editor, tab or terminal | Its own remote machine |
| Who starts each run | You, every time | A schedule, an event, or a standing goal |
| Your role | Watch and steer in real time | Review the result afterwards |
| What it hands back | Answers and suggestions | Finished files, PRs, messages |
| Best for | Ambiguous work where your judgment is the loop | Clear, checkable work you'd rather not babysit |
Most teams need both. The foreground agent helps you think; the background agent does the job you already know how to describe.
Where did background agents come from?
Background agents started in software development. Cursor named its remote agents "background agents" (it now calls them Cloud Agents). Each one, in Cursor's words, "runs on its own dedicated virtual machine" with your repo and dependencies, and opens a pull request when it finishes, so you can "close your laptop and check the result later" (Cursor).
GitHub followed the same pattern with the Copilot coding agent, which picks up an issue and returns a draft PR. Compared with an in-editor assistant, the shift is simple: the work leaves your machine.
Coding went first because code has a built-in review step. A pull request is a finished, checkable artifact — exactly what you want from an agent you weren't watching.
Long-running AI agents in 2026: what shipped in September
Within three weeks, four vendors took background agents out of the IDE and into general work. Each one leads with the same promise: the agent keeps going when you don't.
| Launch | Announced | What wakes it up | Where it runs | Memory across runs | Approval gate | Status |
|---|---|---|---|---|---|---|
| Salesforce Hunter | Sep 11 | A goal, e.g. "rescue at-risk deals before quarter end" | Agentforce long-horizon runtime | Yes — plans persist "across days and weeks" | Guardrails set when seller approval is required | Pilot; GA November 2026 |
| Microsoft Copilot Autopilot | Sep 25 | An objective with boundaries you set | Cloud, in your tenant, with "its own identity, memory, computer and workspace" | Yes | Permissions and audit built in | Private preview (formerly Scout) |
| Perplexity Computer Automations | Sep 29 | A schedule, or events in Slack, Gmail, Outlook, Linear, GitHub | Perplexity Computer | Yes — "memory of prior runs" | Engineering runs open PRs for review | Live for Computer users |
| OpenAI dots | Sep 29 | Goals you assign; looks for work when idle | Its own cloud computer and browser | Learns from your feedback | "Auto-review" before consequential actions | Rolling out to Pro and Business Premium; not yet for Pro users in EEA, Switzerland or UK |
Sources: Salesforce, Microsoft, Perplexity, SiliconANGLE on OpenAI dots.
Microsoft's own line sums up the category: Autopilot "keeps working while you sleep or your attention is elsewhere." Salesforce frames it as durable execution: the plan keeps running and "work doesn't stop when the interaction ends."
Hunter and Autopilot live inside a vendor's suite (Salesforce, Microsoft 365). Dots and Perplexity Automations live inside a chat product. If your team works across several tools, check what the agent can reach before you check how smart it is.
What does a background agent need to keep working?
Five things, and the launches above all have them. Remove any one and the agent either stops when you leave or works blind.
- A trigger. A schedule ("every Monday at 8:00") or an event ("when a Linear ticket gets this label"). Without one, someone still has to remember to start it. Our guide to event-driven AI agents covers the event side.
- Its own computer. A browser, a file system, somewhere to run code. This is what lets it open the dashboard and download the CSV instead of telling you how. Cursor, OpenAI and Microsoft all put the word "computer" or "virtual machine" in their pitch for this reason.
- Memory across runs. The second run should know what the first one found. That's the difference between persistent and ephemeral AI agents — persistent AI agents compare this week with last week; ephemeral ones start from zero every time.
- An approval gate. A rule for which actions it takes alone and which wait for a human: Salesforce's "seller approval," OpenAI's "auto-review."
- A place to report, plus a record. The result has to land where people already look — a Slack channel, a PR, an inbox — with a transcript of what the agent did to get there.
Ask of any background agent: what wakes it up, what machine does it work on, what does it remember from last time, what does it need permission for, and where will I see the result? Five clear answers means it can run unattended.
What still needs you?
Judgment, exceptions and sign-off. Background agents are good at recurring, checkable work: reports, data pulls, monitoring, follow-ups. They are weaker when the task is ambiguous or the stakes are high, and even OpenAI tells dots users to review consequential work because the agents "can still make mistakes."
Three habits keep background agents useful:
- Start with work you can check in a minute. A report with numbers you recognise beats an open-ended "grow our pipeline."
- Read the first few transcripts. The first week is onboarding. Corrections made then are the ones that stick.
- Keep money, customers and anything irreversible behind a human approval — the pattern in our human-in-the-loop guide.
An agent you can't audit is an agent you'll eventually switch off. Pick one that leaves a readable record of every run.
Put a Background Agent on Your Team
The September launches tie background agents to one vendor's suite or chat app. cowork.ink gives you AI coworkers instead, as many as you need in whatever roles you need, and each one has its own cloud computer with a browser, files, an inbox and a code runtime. It signs into your tools through its browser, works on the schedule you set, and posts the finished file in Slack or Telegram. Every run leaves a transcript you can read and correct.
Setup takes about three minutes: create a coworker, tell it what it's for, hand it what a person gets on their first morning, connect Slack or Telegram, and give it the first task in plain language. One flat price covers every coworker, their machines and the models they use (Claude, DeepSeek, Gemini and other frontier models, routed per task): $25/month billed yearly, or $29 monthly, with a 15-day money-back guarantee. For more on agents that sit inside the team, read the case for coworker AI.