Agent rollout decision record

Google Jules

Conditional pilot Use Jules on selected GitHub repos and low-risk bugs; require scoped installation, branch protection, CI, and human review.

Check this agent for your environment Compare alternatives Methodology

Verdict

Conditional pilot

Async GitHub coding tasks are pilotable only after repo-scope and branch/PR review boundaries are explicit.

Decision depth: AI-assisted launch evidence record

Evidence audit: AI-assisted evidence audit

Permission footprint

Repository readnot evidenced

File editsexplicitly evidenced

Shell commandsnot evidenced

Pull requestsnot evidenced

Issue/PR commentsexplicitly evidenced

Network/API callsnot evidenced

Secrets/env varsnot evidenced

Sandboxingexplicitly evidenced

Permission controlsnot evidenced

Audit loggingnot evidenced

Enterprise adminnot evidenced

Branch isolationnot evidenced

Human approvalexplicitly evidenced

Production deploymentnot evidenced

Power level

Level 3: executes actions

Highest-risk evidenced power: pull-request creation

Controls separated

Required before pilot — organization must implement

least-privilege permissions, isolated branch, protected branches, mandatory human review, sandbox, command allowlist, audit logs, separate test credentials

Vendor-native control evidenced

branch isolation, security docs, admin console

Vendor-native control not evidenced

native sandbox, permission system, audit logs, SSO, data-retention settings, approval workflow

Pilot plan

Task class: Terminal and coding agent

Environment: non-production repository

Maximum access: pull-request creation

Reviewer: engineering manager plus security reviewer

Success criteria: small reversible code changes pass tests and review without secret exposure

Stop conditions: unexpected command execution, broad repo access, failed audit trail, or reviewer cannot understand diffs

Prohibited: production deploys, unreviewed merges, secret access, large autonomous rewrites

Blocking issues and unknowns

Blockers: reads repository is not explicitly evidenced in the current source set.

Unknowns: Current availability and repo permission details require verification.

Limitations: Async execution obscures local command review unless logs are available.

Claim-level evidence

edits files

Google Jules: edits files

Status: explicitly-supported · Reviewer: AI-assisted evidence audit · Confidence: 86/100

Source: Getting started (official-docs)

Excerpt: Jules runs in a virtual machine where it clones your code, installs dependencies, and modifies files.

Limitations: Official-source excerpt is claim-specific; tenant/workspace configuration can still change rollout risk.

supports sandboxing

Google Jules: supports sandboxing

Status: explicitly-supported · Reviewer: AI-assisted evidence audit · Confidence: 86/100

Source: Getting started (official-docs)

Excerpt: Jules runs in a virtual machine where it clones your code, installs dependencies, and modifies files.

Limitations: Official-source excerpt is claim-specific; tenant/workspace configuration can still change rollout risk.

supports human approval controls

Google Jules: supports human approval controls

Status: explicitly-supported · Reviewer: AI-assisted evidence audit · Confidence: 86/100

Source: Getting started (official-docs)

Excerpt: Once you submit a task, Jules will generate a plan. You can review and approve it before any code changes are made.

Limitations: Official-source excerpt is claim-specific; tenant/workspace configuration can still change rollout risk.

comments on issues or PRs

Google Jules: comments on issues or PRs

Status: explicitly-supported · Reviewer: AI-assisted evidence audit · Confidence: 86/100

Source: Running Tasks with Jules (official-docs)

Excerpt: Soon, you will see Jules comment automatically on the issue. When Jules is finished with the issue, it will provide a link to the pull request where you can review its work.

Limitations: Official-source excerpt is claim-specific; tenant/workspace configuration can still change rollout risk.