Jules
Autonomous coding agent from Google
About Jules
Jules is an autonomous coding agent from Google Labs that works on a repository rather than in your editor. You give it a GitHub repo and a description of what you want, and it goes away and does the work: it clones the code to a cloud virtual machine, forms a plan, makes the change, and opens a pull request for you to review.
The design choice worth understanding is that it is asynchronous. Most AI coding tools are synchronous, sitting in your IDE and responding while you wait. Jules is closer to delegating a ticket. You describe the job, close the tab, and come back to a diff. That suits a different class of work: the version bump nobody wants to do, the test coverage gap, the small bug with a clear reproduction, the dependency upgrade that touches thirty files mechanically.
It runs on Gemini, with the free tier using Gemini 2.5 Pro and paid tiers Gemini 3 Pro. Because the work happens on a Google-run VM rather than your machine, it can take as long as it needs without blocking you, and several tasks can run at once. The free tier allows 15 tasks a day with 3 running concurrently, which is enough to find out whether it fits your workflow before paying anything.
You connect a GitHub repository and write a prompt. The more specific the prompt, the better this works, in the same way a well-written ticket produces better results from a human.
Jules clones the repository to a cloud VM. Working on a real checkout rather than on file snippets means it can run the build, run the tests, and see what actually breaks.
It then produces a plan before changing anything, which you can inspect. Once it works, it generates a diff and opens a pull request against your repo.
You review that pull request exactly as you would a colleague's. Nothing merges without your approval, and the review is the control: the agent is trusted to propose, never to land.
Concurrency is the throughput lever. Three tasks at once on the free tier, up to sixty on Ultra, so a backlog of small mechanical jobs can be worked through in parallel rather than one at a time.
- •Asynchronous by design - delegate a task and leave, rather than supervising a chat. The unit of work is a pull request, not a conversation.
- •Real repository checkout - it clones to a VM and can build and test, so it sees genuine failures instead of guessing from context snippets.
- •Plan before diff - the intended approach is visible before code changes, which is where a misunderstanding is cheapest to correct.
- •Pull request output - work arrives through the review process you already have, so nothing bypasses your existing controls.
- •Parallel tasks - 3 concurrent on the free tier, 15 on Pro, 60 on Ultra, which turns a queue of chores into a batch.
- •Free tier that is actually usable - 15 tasks a day is enough for real evaluation, not a demo.
Maintainers with a backlog of unglamorous work benefit most. Dependency updates, flaky test fixes, small well-specified bugs and mechanical refactors are exactly the jobs that are easy to describe, tedious to do, and safe to review.
Teams that already work through pull requests get the tightest fit, because the output lands in the process they already trust rather than asking them to adopt a new one.
Anyone evaluating agentic coding cheaply should look here, since the free tier is generous enough to reach a real conclusion.
It is a poor fit for exploratory work where you do not yet know what you want. Asynchronous delegation needs a well-formed request; if you are still thinking, an interactive assistant in your editor will serve you better.
Pricing
From $0/mo
Free tier gives 15 tasks a day and 3 concurrent tasks. Jules Pro raises that to 100 tasks a day and 15 concurrent with Gemini 3 Pro, and Jules Ultra to 300 tasks a day and 60 concurrent with priority access to new models. Paid tiers are sold through Google AI subscriptions rather than priced on the Jules page.
From the vendor pricing page, 2026-09-22
