Quickstart

Set up Cloud Agents for a repository, configure safety, and launch your first review and execute runs.

This guide gets a repository from zero to its first two useful workflows:

  1. a review run that inspects an incoming change, and
  2. an execute run that prepares a scoped code change under approval.

Prerequisites

  • A Skytells account
  • Pro plan or higher — see Account Plans
  • A GitHub repository you are authorized to connect
  • A repository owner who can decide approval and protected-path policy

Before You Turn It On

Before a repository is connected, decide a few things explicitly.

DecisionWhy it matters
Who owns repository policySomeone needs authority to decide what is protected, ignored, or eligible for automation.
Who approves sensitive outputsApproval queues fail quickly when the product is configured but nobody owns the decision.
Which branch flow mattersRuns are only useful if they operate on the same branch model your team actually uses.
What validation is trustedValidate after execute only helps when CI or validation signals are already meaningful.
Which paths are sensitiveProtected paths should be named before execute workflows are trusted on production repositories.

This preparation is the difference between a controlled rollout and a noisy one.

Step 1: Open the Cloud Agents Workspace

Confirm plan access

If Cloud Agents is not available in your workspace, verify that the account is on Pro or above and that the feature is enabled for the workspace.

Review the top-level tabs

The workspace opens with tabs for Overview, Chat, Runs, Usage, Repositories, Approvals, Add-ons, and Settings.

Step 2: Connect GitHub and Verify Repository Access

Open Settings

Go to Settings inside Cloud Agents.

Check GitHub integration

In the settings overview, confirm that GitHub Integration shows a connected state and that the repositories you expect are listed.

Manage repository authorization if needed

If the repository is missing, use Manage Repositories and complete the GitHub authorization changes described in GitHub Integration.

Step 3: Add or Select a Repository

Open Repositories

Go to Repositories in the Cloud Agents workspace.

Add the repository

Use Add repository if the repository is not already configured.

Select the repository

In the Repository List panel, select the repository you want to automate.

Inspect defaults

Review the repository settings card group, especially Auto review, Auto execute, Review on PR events, and Validate after execute.

Step 4: Set Safety Before You Dispatch

Use conservative settings first.

ControlRecommended first valueWhy
Auto reviewEnabledSafe first pass for incoming change evaluation.
Auto executeDisabled or limitedPrevents unattended write-style workflows while rules are still new.
Review on PR eventsEnabledEnsures PR activity produces timely analysis.
Validate after executeEnabledAdds a verification step before risky changes are considered complete.
Protected pathsAdd critical directoriesKeeps automation away from high-risk files until you trust the workflow.
Ignore pathsAdd generated contentReduces noisy reviews and false-positive changes.

See Repositories for the full repository policy model.

Choose a Rollout Style

Different teams should not start the same way.

StyleBest forRecommended setup
Observer modeTeams new to the product or new to repo automationManual review runs only, execute disabled, protected paths defined early
Review-first modeMost teams adopting Cloud Agents for the first timeAuto review on, PR signals on, execute manual, approvals active
Controlled execute modeTeams that already trust the repo policy and want more throughputExecute allowed in narrow scopes, validation on, approval checkpoints retained

If you are unsure, choose Review-first mode. It gives the team value quickly without treating write-style automation as the starting point.

Step 5: Choose an Agent

Cloud Agents ships with system agents for core review and execute flows, and supports custom agents for team-specific behavior.

Start with a system agent

For the first pass, choose the built-in Code Review Agent for review runs and Execute Agent for code-change runs.

Use a custom agent only when the repository needs it

Add a custom agent once you know what should be specialized: migrations, docs, refactors, release checks, or another consistent operating pattern.

See Agents for agent design guidance.

Step 6: Launch Your First Review Run

Open Chat

Go to Chat.

Scope the run

Select the repository, branch, and agent in the top controls.

Choose the review path

Start a New Review run or select Code Review inside the new-run flow.

Write a precise task

Example:

Review the open pull request for auth callback changes. Focus on routing regressions, unsafe redirects, and missing tests. Do not suggest style-only edits.

Submit and inspect

Track the run in Runs. Review the timeline, findings, and any approval requirement that appears.

Step 7: Launch Your First Execute Run

Start a new execute flow

From Chat or Overview, choose New Execute.

Select execute mode

In the new-run setup, choose Execute Change.

Keep scope narrow

Use repository and branch selectors, then provide a task with a clear goal and a small blast radius.

Example:

Add empty-state messaging to the approvals table in the dashboard. Limit changes to the approvals page and related tests.

Review plan or direct execution mode

Use Plan first if the repository is new to Cloud Agents. Move to Agent mode once the operating pattern is stable.

Approve only after inspection

If the run pauses for approval, inspect the run detail page before approving or rejecting it.

Step 8: Confirm the End-to-End Loop

After the first review and execute runs complete, verify the following:

  • The repository appears under Repositories with the expected automation rules.
  • The run history is visible under Runs.
  • Any gated step appears under Approvals.
  • Consumption appears under Usage.

What Success Looks Like After Week One

  • Time to first review is materially lower than before.
  • The approval queue is active but not overloaded.
  • Most runs are Completed or intentionally Awaiting, not failing for avoidable setup reasons.
  • Repository rules have already been refined from real run data instead of staying at their defaults.

If the First Week Goes Poorly

SymptomUsually means
Too many noisy findingsPrompts are too broad or ignore paths are incomplete
Too many approval pausesProtected paths or execute scope need refinement
Too many failuresBranch scope, validation assumptions, or agent choice is wrong
Nobody trusts the outputThe team started too aggressively; move back to review-first and tighten ownership

Next Steps

How is this guide?

On this page