Jules Async Coding Agent 2026 — VM-Powered “Wait-Free AI”

“I asked Gemini CLI to refactor the whole repo and my terminal was tied up for thirty minutes.” “Claude Code is still working on test fixes; I went to make coffee and it is still going when I come back.” This fifth article in the series unpacks the Google-built async agent that structurally erases the time you spend waiting on AI.

Gemini CLI is a synchronous terminal agent — the user sits with the AI and watches the work happen. Jules inverts that model. Tasks run inside isolated Google Cloud VMs while you continue with other work, and you review the result when notified. Run multiple tasks in parallel; compress the wait time to zero is its core value proposition.

Synchronous vs Asynchronous — Two Worldviews

AI coding agents split into two camps. Synchronous (Cursor, Claude Code, Gemini CLI, Codex CLI) sits the user at the editor or terminal, engaged turn by turn. It demands attention but offers immediacy and a strong sense of control. Asynchronous (Jules, Cognition Devin, GitHub Copilot Workspace) takes a task, runs it in the background, and notifies you when it is done. You trade immediacy for parallelism.

The two camps are complements, not competitors. Short, well-scoped bug fixes fit the synchronous model. Repo-wide refactors, multi-file API migrations, and large documentation passes fit asynchronous. Mature workflows use both.

Why Jules — Three Specific Strengths

1. Bundled with Google AI Pro

Jules is included with the $19.99/month Google AI Pro subscription. Devin starts at $500/month for the Team tier; GitHub Copilot Workspace is bundled with Copilot but the workspace tier is metered. For individual makers, the cost-of-entry difference is decisive.

2. Secure VM × Direct GitHub Integration

Every Jules task runs in a fresh Google Cloud VM. Your local credentials never leave your machine; the agent never touches your filesystem. Output arrives as a GitHub pull request, ready for code review. The security model is unusually clean: nothing the agent does can affect your local environment without your explicit merge.

3. Plan-Before-Execute Transparency

Jules presents its plan before touching code. You can edit the plan, narrow the scope, or cancel. During execution, intermediate steps stream to the UI. After execution you can request revisions before the PR is finalised. This avoids the failure mode of “agent runs for 40 minutes, produces unusable output, you have no idea where it went wrong.”

The Four-Step Execution Model

• Task submission: from jules.google’s web UI, the Jules Tools CLI, the Jules API, or the Gemini CLI extension.
• VM provisioning: a secure Google Cloud VM boots, the GitHub repo is cloned, dependencies installed.
• Plan and execute: Jules presents its plan, you approve, the agent edits code with Gemini 2.5 Pro as the underlying model (with migration to Gemini 3.x expected).
• PR creation: a labelled PR appears in your repo with a structured description, test results, and a diff for review.