Jules

What Jules is

Jules is an asynchronous coding agent from Google. The word "asynchronous" describes how the developer hands off work: instead of staying in the editor while a model suggests lines as they type, the developer gives Jules a task and comes back later to inspect the result. Google first previewed it on 11 December 2024 as an experimental code agent that integrates with GitHub workflows and can take on an issue, draft a plan and execute that plan under developer supervision^[1].

That description sets out the basic loop the product is built around: a task goes in, a plan comes out, work is carried out, and a human checks it. The supervision step is not incidental. Jules is designed to hand back its changes as something a developer reviews rather than as code that lands silently in a repository.

From preview to open beta

After the December preview, Google widened access on 20 May 2025, opening a public beta to everyone with no waitlist, in every place where the Gemini model was available^[2]. Removing the waitlist changed the product from a limited experiment into something any developer in a supported region could try.

During that beta, Jules ran on Gemini 2.5 Pro and was free to use with usage limits; Google said pricing was expected later^[2].

How it works

When given a task, Jules clones a codebase into a secure Google Cloud virtual machine and works there. Within that environment it can write tests, build features, provide audio changelogs, fix bugs and bump dependency versions, and it presents a diff of the changes it has made^[2].

The list is worth reading closely, because it mixes two kinds of work. Some of it is new code — building features, writing tests. Some of it is upkeep: fixing bugs and raising the version numbers of the third-party libraries a project depends on, the sort of repetitive maintenance that accumulates in any long-lived repository. The audio changelog is a way of summarising what happened in a run without reading it all.

Google also addressed a question that comes up whenever a hosted coding tool touches private code: the company said Jules does not train on users' private code and keeps data isolated within the execution environment^[2]. The diff itself keeps the developer in the reviewing position — the agent proposes, the person disposes.

General availability and limits

SiliconANGLE reported on 6 August 2025 that Google launched Jules in general availability^[3]. At that point the free version allowed 15 individual daily tasks and three concurrent tasks, and the agent was still powered by Gemini 2.5 Pro^[3].

Those numbers say something about what an asynchronous agent costs to run. Each task is not a keystroke but a session in a virtual machine that clones a repository and works through a plan, so usage is counted in tasks rather than in tokens or minutes. The cap on concurrent tasks matters for the same reason: several runs can be in flight at once, and the limit decides how much parallel work a single account can set going.

Where it fits

Jules belongs to a category of coding tools that changed shape once models became reliable enough to work for minutes at a time without a human watching every step. In-editor assistants optimise for the moment of typing; asynchronous agents optimise for the task that can be described in an issue and picked up later. The tasks Jules was built for — writing tests, fixing a bug, bumping dependency versions, building a described feature — are the ones that fit that handoff well: they are stated in words, they end in a diff, and the diff can be judged on its own.

Because the work happens away from the editor and comes back for review, the product's practical value depends on how well the plan and the diff communicate what was done. Jules was designed around exactly those two artefacts, from the December 2024 preview onward^[1][2].