Background AI Agents: What They Are and How They Work

A background AI agent is a coding agent that runs on someone else's machine while you work on something else. You hand it a task from your editor or an issue tracker. It comes back later with a branch or a pull request. Five major vendors ship one now, and most have renamed theirs at least once. Each run starts from a fresh clone. What it remembers from the last run depends on files in your repository and on vendor features that stay inside one product.

What is a background AI agent, in one sentence?

A background AI agent is a coding agent that works unattended in a cloud VM with a copy of your repository, then returns a pull request.

Cursor shipped one of the first, as a preview in May 2025. GitHub's coding agent and Google's Jules followed within a week. Anthropic added Claude Code on the web on October 20, 2025. Since then the names have shifted. Cursor's docs say Cloud Agents were formerly called Background Agents, and GitHub's docs now say Copilot cloud agent.

How the main background agents compare

The last column decides how much you re-explain to the agent on each run. Sources are linked in the sections below.

How does a background agent run?

The pattern is the same across vendors. The agent gets a fresh VM and clones the repository. After a setup step installs dependencies, it works through the task and pushes a branch. Claude Code's cloud environment docs describe an Ubuntu 24.04 VM with 4 vCPUs and 16 GB of RAM. They also say cloud sessions start from a fresh clone of your repository.

Setup is the slow part, so every vendor caches it. Most of them snapshot the environment after setup, Jules included, per its environment docs. Claude Code keeps its snapshot for roughly seven days. Codex's cloud environments keep saved state for up to seven days after your last turn.

Network access is restricted by default. Codex blocks internet access during the agent phase but lets setup scripts reach the internet. Copilot runs behind a firewall that limits the agent's internet access. The reason is prompt injection from untrusted web pages, which OpenAI names in its docs.

What can background agents trigger on?

Most start from a person. Cursor's Automations, launched on March 5, 2026, can start on a schedule or from events such as a webhook. Jules added scheduled tasks in December 2025, per its changelog. Claude Code cloud sessions can run from scheduled or triggered routines.

GitHub's agent starts from an issue: you assign it one and it opens a draft pull request. Its docs cap a session on Actions at 59 minutes. It can only push to its own branch, and only users with write access can trigger it.

What does a background agent remember between runs?

Very little by default. Each run starts from a fresh clone, so the main channel for context is files in the repository. As Cursor's rules docs put it, large language models don't retain memory between completions. AGENTS.md is the shared format. All five products read it, Claude Code in place of CLAUDE.md.

Personal context does not travel. Claude Code's cloud docs say the repository's CLAUDE.md comes along with the clone. Your user-level CLAUDE.md stays on your machine. Its memory docs add that auto memory files are not shared across machines or cloud environments. Codex says personal skills from your local computer are not synced to cloud environments.

Some vendors added their own memory, each scoped to its own product:

• Copilot Memory stores facts per repository. It checks each fact's citations against the current branch before use and expires facts after 90 days by default.
• Cursor Automations have a memory tool for persistent notes across runs of the same automation. Cursor warns to use it with caution on untrusted input.
• Jules added memory on September 30, 2025, which learns from your interactions per repository.

A fact Copilot learned is invisible to Jules.

Read the fine print

Every capability above comes from vendor docs, and we found no independent measurements of how often runs succeed. Codex's older container-based cloud is now labeled legacy and slated for deprecation. Its 12-hour cache does not apply to the newer VM-based environments. GitHub's billing moved from Actions minutes plus premium requests to an account allowance during 2026. Copilot Memory launched in public preview in January 2026, and the current docs no longer state its status.

Cursor bills cloud agents against your plan's usage. Claude Code says there is no separate charge for the cloud VM, but sessions share rate limits with your other Claude usage. Jules counts tasks per rolling 24 hours, with limits set by plan.

The bottom line

Background agents work well for tasks you can describe in an issue and verify in a pull request. Put build commands and conventions in AGENTS.md, because it is the one file every product reads. Do not count on personal or vendor-specific memory following a task into the cloud.

How do you give background agents the same memory as your local agent?

User-level CLAUDE.md and auto memory stay on your machine, so a Claude Code cloud session starts without them. Each vendor's own memory is locked to that vendor. Often it is locked to one repository too. A team that uses Cursor for automations and Claude Code for reviews keeps two separate memories of the same codebase.

memU keeps one shared memory that several agents and devices can use. Its host adapters cover Claude Code and Codex, among others. A background task slices new session logs into jobs, and the agent distills them into memory and skill files in readable Markdown. memU makes no LLM calls of its own.

A standing instruction in CLAUDE.md or AGENTS.md tells the agent to retrieve before answering. Support varies by host and mode, so check the support matrix before relying on it for a given cloud product. The design is explained in why memU leaves the thinking to the agent.

To try the hosted version, which works across devices, get an API key from memu.so and send your agent one message:

Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU, API Key is YOUR_MEMU_API_KEY.

The support matrix and self-hosted setup are in the memU GitHub repository.

FAQ

What are Cursor background agents called now?

Cloud Agents. Cursor's docs say Cloud Agents were formerly called Background Agents. They run in isolated cloud VMs and produce pull requests, and Cursor Automations can start them on a schedule or an event.

Do background agents have internet access?

Limited by default. Codex blocks internet access during the agent phase but allows it for setup scripts, and Copilot's agent runs behind a firewall you can configure.

Does a background agent remember previous tasks?

Only through files and vendor features. Each run starts from a fresh clone, so AGENTS.md and other repository files are the main memory. Some vendors add memory features, each scoped to their own product.

Is Claude Code on the web the same as a background agent?

Yes. Claude Code cloud sessions run in an isolated, Anthropic-managed VM. They work from a clone of your repository and can create pull requests without your machine.

How long can GitHub Copilot's cloud agent run?

A session running on GitHub Actions has a maximum execution time of 59 minutes, which cannot be extended.