Jules by Google
Experimental coding agent from Google that clones a repo to a secure cloud VM plans a change with Gemini executes edits runs tests and opens a review so you supervise reliable PRs end to end.
Overview
Jules is an experimental coding agent from Google designed to handle scoped engineering tasks with human oversight. You label a GitHub issue or call the CLI then Jules fetches your repository to a secure Google cloud VM. It drafts a concrete implementation plan using the latest Gemini models shows the exact files it will touch and asks for approval.
When you confirm it edits code runs dependency install and tests and validates the result against your acceptance criteria. The agent then opens a pull request with a summary of decisions and links to diffs so reviewers move faster. Because Jules lives close to your repo and CI you avoid heavy local setup.
Teams use it for framework upgrades mechanical refactors dependency bumps documentation fixes and small features that are well specified. For critical changes you stay in the loop confirm steps and enforce code owners. Early access materials emphasize auditability safety rails and repeatable workflows rather than black box magic which makes it attractive for production teams that need rigor with speed.
Key features
- Issue to PR workflow with explicit plan steps and file lists that you approve before execution and merge
- Secure cloud VM per task so no local setup and a clean environment with ephemeral resources and logs
- Deep repo understanding via Gemini planning that maps tasks to files and tests with clear acceptance checks
- Automated edits runs and test execution with visible output so reviewers trust the proposed changes
- Pull request creation with structured summary rationale and diffs to streamline team review flows
- Scoped permissions using repo tokens and granular access so risk is minimized during automated work
- CLI and label based triggers so you can start tasks from terminal or GitHub without custom glue code
- Human in the loop checkpoints at plan approve and PR stages so engineering standards stay intact
Best for
- Upgrade framework versions across services with reproducible steps and validation evidence for reviewers
- Apply mechanical refactors at scale such as path changes or API shifts while preserving behavior with tests
- Fix flaky test suites by instrumenting runs and proposing targeted stabilizations that ship quickly
- Generate missing documentation and examples that match code reality to reduce onboarding time
- Patch security alerts by bumping dependencies and running checks to validate the supply chain change
- Create scaffolds for small features based on an issue template that encodes acceptance criteria
- Backport fixes to a maintenance branch with a minimal diff and CI verification for compliance
- Prepare migration PRs that split changes into stages so risk and reviewer load remain manageable
Capabilities
Gemini Task Planning
Jules proposes a concrete change plan maps files to edits and aligns on acceptance checks so scope is crisp and risks are visible before any code moves.
Secure VM Runs
Each task runs in a clean cloud VM with installs tests and logs captured which keeps local machines untouched and results reproducible.
Explained Pull Requests
The agent opens a PR with summaries and diffs so reviewers see what changed and why which reduces back and forth and accelerates merges.
Human Approvals
You approve the plan gate execution and enforce code owners so automated changes stay compliant with team processes and standards.
Frequently Asked Questions
Is Jules available to the public and what does it cost?
Jules is currently offered as an experimental Google experience with sign in access and no published list price so treat availability and features as subject to change and verify terms before relying on it.
How does Jules keep my repository secure during automated work?
Tasks run in ephemeral cloud VMs with scoped repository tokens and logs you can audit which limits blast radius while maintaining traceability of edits and test output.
Can I trust the plan before code is changed?
Yes the workflow is plan first with explicit files and steps then your approval gates execution which avoids surprise edits and keeps reviewers aligned.
Does Jules run my test suite and show results inside the PR?
The agent installs dependencies runs tests and attaches logs and summaries to the PR so reviewers can validate outcomes without re running everything locally.
What size of tasks are best suited for Jules today?
Framework upgrades mechanical refactors doc fixes and well specified features are ideal while multi repo architecture changes still need human orchestration first
