The call
Make passing CI checks the agent's condition for asking for review, and keep merge approval and deploys with people.
Set a monthly spend limit for Claude Code Review before you turn it on, because it bills tokens through usage credits with no flat per-review price.
Claude Code needs no new planning or review tools. Sanity, Expo, incident.io, Spotify and Fin (formerly Intercom) run it inside the tools they already had, and people keep merge approval and deploys. It reads the issue from your tracker, works in its own git worktree, runs the checks CI runs, and hands a pull request to your reviewers.
When you need this
- Engineers copy issue text into a terminal and paste results back by hand.
- Parallel agent sessions collide in one working copy.
- Agent pull requests reach reviewers without the context they need.
Worked example: from ticket to merged change at four teams
Plan
Sanity connects Linear, Notion and GitHub as context, and Expo found the Linear, Sentry, Figma and Graphite MCP servers useful for turning an issue into a brief.
Implement
incident.io runs four or five agents at once, each started with claude --worktree in its own git worktree on its own branch.
Test
Spotify's background agent checks its own work through a verify tool that runs formatters, linters and tests and summarizes the logs.
Open the pull request
Intercom blocks a raw gh pr create until its create-pr skill has run, so agent pull requests follow the team's conventions.
Issue-to-pull-request workflow map
Plan: from issue to brief
Where does the task come from?
Where Claude Code connectsKeep the issue in your tracker and give the agent access through the tracker's CLI or an MCP server. Anthropic notes that CLI tools such as gh use less context than MCP servers.
Implement: one task per isolated checkout
How do parallel sessions avoid colliding?
Where Claude Code connectsThe command claude --worktree <name> starts a session in its own git worktree on its own branch. For large mechanical changes, /batch splits the work across 5 to 30 subagents, each in its own worktree.
Test: the checks CI will run
Can the agent run the same checks before it asks for review?
Where Claude Code connectsList the commands in CLAUDE.md, and use a Stop hook to run them before a turn can end.
Open the pull request
What does a reviewer see?
Where Claude Code connectsWrite your pull request conventions into CLAUDE.md or a skill, and use a PreToolUse hook to block a raw gh pr create until that skill has run.
Review: people decide, agents assist
Which review comments come from an agent, and who approves the merge?
Where Claude Code connectsClaude Code Review posts severity-ranked inline comments. Set its monthly spend limit and its review trigger before it runs, and use a REVIEW.md file to tune what it flags. Human approval stays the merge gate.
Start work from GitHub, GitLab, and Slack
Which work can start without a terminal?
Where Claude Code connectsClaude Code's GitHub Action answers @claude mentions in issues and pull requests or runs a prompt on a schedule. Only users with write access can trigger it. GitLab CI/CD support is in beta and maintained by GitLab. On Team and Enterprise plans, Slack requests go through Claude Tag, which runs @Claude as the organization's shared identity.
Questions to settle with each team
- What must a ticket contain before an agent can take it?
- Who may trigger @claude in CI, and what turn limit (--max-turns) and workflow timeout does each run get?
- Does Claude Code Review run once per pull request, on every push, or only on request, and what is its monthly spend cap?
- How do reviewers know a change came from an agent session?
- Which steps stay with people, such as merge approval and production deploys?
Red flags
- An agent holding deploy credentials in CI.
- Several sessions editing one working copy.
- Review on every push with no monthly spend cap for Code Review.
- A separate tracker or planning tool created only for agent work.
Products in this piece
Sources
- Official documentation
- 9
- Public source
- 5
- Claude Code docs: Common workflows
- Claude Code docs: Best practices
- Claude Code docs: Manage costs effectively
- Claude Code docs: Hooks
- Claude Code docs: GitHub Actions
- Claude Code docs: GitLab CI/CD
- Claude Code docs: Claude Code in Slack
- Claude Code docs: Code Review
- Claude Code docs: Code Review costs and usage
- incident.io: How we're shipping faster with Claude Code and Git Worktrees (Jun 27, 2025)
- Intercom: How we use Claude Code today at Intercom (Brian Scanlan, Mar 19, 2026)
- Spotify Engineering: Context Engineering for Background Coding Agents, Part 2 (Nov 24, 2025)
- Sanity: First attempt will be 95% garbage (Vincent Quigley, Sep 2, 2025)
- Expo: What our web team learned using Claude Code for a month (Juwan Wheatley, Feb 6, 2026)