Cursor Automations Developer Guide: Always-On AI Coding Agents

Last updated: June 15, 2026

Cursor Automations shipped in March 2026 and changed how AI coding agents fit into development workflows. Instead of waiting for you to prompt them, automations let agents run in the background based on triggers: test failures, file saves, PR opens, cron schedules, or webhook calls.

The shift is from interactive AI to always-on AI. Teams using automations report 20-40% reductions in manual review tasks and 1-2 hours recovered per developer weekly. Here is how to set them up.

What Automations Actually Do#

Traditional AI coding assistants are reactive. You ask a question, they answer. You prompt a refactor, they generate code. The loop requires you to be present and typing.

Automations flip this. You define a trigger and instructions once. The agent runs whenever that trigger fires, whether you are at your desk or not. Results queue up for review.

The practical difference:

Jonas Nelle, Cursor's engineering lead for async agents, described it as a "conveyor belt" for development: "Humans are not completely out of the picture. Instead, they are not always initiating. They're called in at the right points."

Setting Up Your First Automation#

Automations are configured via YAML files in your project's .cursor/automations/ directory, or through Cursor's settings panel.

Directory Structure#

my-project/
  .cursor/
    automations/
      fix-tests.yaml
      dependency-audit.yaml
      pr-review.yaml

Basic Configuration#

Every automation needs three things: a name, a trigger, and instructions.

name: Fix Failing Tests
trigger:
  - type: test_failure
    test_command: pnpm test
    debounce_ms: 5000
instructions: |
  Review failing tests and source code. Identify root cause
  and propose minimal fix without modifying test expectations.
  Run tests in sandbox to verify before submitting diff.
sandbox:
  install_command: pnpm install
  env:
    NODE_ENV: test

When your test suite fails, this automation spawns an isolated container, clones your repo, runs the agent with your instructions, and stages a diff for your review.

Trigger Types#

Cursor supports six trigger types. Each fits different workflow patterns.

1. Cron Schedule#

Standard cron syntax for recurring tasks.

name: Weekly Dependency Audit
trigger:
  - type: cron
    schedule: "0 9 * * 1"  # Monday 9am
instructions: |
  Run npm audit. For vulnerabilities with severity high or critical,
  check if patch versions are available. Apply patches and run tests.
  Skip major version bumps.

Good for: dependency updates, dead code detection, documentation sync, weekly summaries.

2. Git Events#

Triggered on push, commit, PR open, PR merge, or branch creation.

name: PR Security Review
trigger:
  - type: git
    event: pull_request_open
instructions: |
  Review the PR diff for security issues. Check for:
  - Hardcoded credentials or API keys
  - SQL injection vectors
  - Unvalidated user input
  - Missing authentication checks
  Post findings as PR comment.

Good for: automated code review, security scanning, style checks.

3. Test Failure#

Fires when your test command exits non-zero.

name: Auto-Fix TypeScript Errors
trigger:
  - type: test_failure
    test_command: pnpm typecheck
    debounce_ms: 10000
instructions: |
  Read the type errors. Fix them without changing business logic.
  Prefer narrowing types over adding type assertions.
  Run typecheck in sandbox to verify fix.

Good for: type errors, linting failures, test fixes.

4. File Save#

Immediate response when you save specific files.

name: Update Changelog
trigger:
  - type: file_save
    patterns:
      - "src/api/**/*.ts"
    debounce_ms: 60000
instructions: |
  If the saved file contains API endpoint changes,
  update CHANGELOG.md with a brief description of the change.
  Follow existing changelog format.

Debouncing is important here. 30-60 seconds prevents trigger spam during active editing.

5. Webhook/API#

External systems invoke automations via REST.

name: Incident Responder
trigger:
  - type: webhook
    path: /incident
instructions: |
  Received alert from PagerDuty. Query recent logs,
  identify potential root cause, and draft initial
  incident report with relevant code references.

Invocation:

curl -X POST \
  -H "Authorization: Bearer $CURSOR_API_KEY" \
  https://api.cursor.sh/automations/{id}/trigger \
  -d '{"alert_id": "12345"}'

Good for: CI/CD integration, incident response, external tool hooks.

6. Manual#

Direct invocation through API when you want on-demand runs.

name: Generate Release Notes
trigger:
  - type: manual
instructions: |
  Compare current main branch to last release tag.
  Generate release notes covering new features, bug fixes,
  and breaking changes. Format for CHANGELOG.md.

Agent Execution Model#

Understanding how automations run helps you write better instructions.

Isolated Containers#

Each run spawns a fresh container with:

• A clone of your repository at HEAD
• Your specified install command executed
• Environment variables from vault (not version control)
• Network access to allowlisted domains only
• No access to your local filesystem outside the repo

The agent works in this sandbox. Changes are staged as a diff for your review. By default, nothing auto-applies.

Execution Flow#

1. Trigger fires
2. Container spawned with repo clone
3. Install command runs
4. Agent receives instructions
5. Agent performs work in sandbox
6. Diff staged for review
7. Developer approves before application

Constraints#

• 30-minute execution time limit per run
• No real-time mid-run guidance
• Monthly run quotas on Pro/Business plans
• Beta API subject to schema changes

Production Workflows#

These patterns are what teams actually use.

Daily Dependency Audit#

name: Dependency Security Check
trigger:
  - type: cron
    schedule: "0 8 * * *"
instructions: |
  Run npm audit. For each vulnerability:
  - If patch version available, update and test
  - If minor version available and changelog looks safe, update and test
  - Skip major versions, log them for human review
  Commit passing updates. Report blocked updates.
sandbox:
  install_command: npm ci

TypeScript Error Cleanup#

name: Fix Type Errors on Save
trigger:
  - type: file_save
    patterns: ["**/*.ts", "**/*.tsx"]
    debounce_ms: 30000
instructions: |
  Run tsc --noEmit. If errors exist in saved files:
  - Fix type errors without changing runtime behavior
  - Prefer type narrowing over assertions
  - Do not modify test files
  Stage fixes for review.
sandbox:
  install_command: pnpm install

PR Review Preparation#

name: Pre-Review Analysis
trigger:
  - type: git
    event: pull_request_open
instructions: |
  Analyze the PR diff for:
  - Functions over 50 lines (suggest extraction)
  - Missing error handling
  - Unused imports or variables
  - Test coverage gaps
  Post analysis as PR comment, not blocking.

Dead Code Detection#

name: Weekly Dead Code Sweep
trigger:
  - type: cron
    schedule: "0 10 * * 5"  # Friday 10am
instructions: |
  Identify unused exports, unreachable code paths,
  and files with zero imports. Generate removal PR
  with tests passing. Group related removals.

CI/CD Integration#

Automations work alongside your existing CI. The common pattern is AI-generated insights posted to PRs before human review.

GitHub Actions Example#

# .github/workflows/cursor-analyze.yml
name: Cursor Analysis
on:
  pull_request:
    types: [opened, synchronize]

jobs:
  analyze:
    runs-on: ubuntu-latest
    steps:
      - name: Trigger Cursor Automation
        run: |
          curl -X POST \
            -H "Authorization: Bearer ${{ secrets.CURSOR_API_KEY }}" \
            https://api.cursor.sh/automations/pr-analysis/trigger \
            -d '{"pr_number": "${{ github.event.number }}"}'

This augments human reviewers rather than replacing them.

Configuration Best Practices#

Start Narrow#

Begin with verifiable, low-risk automations: test fixes, dependency updates, dead code removal. These have clear success criteria.

Disable Auto-Apply Initially#

Run with auto-apply off for your first 30 days. Review every diff. Build trust in the agent's judgment before expanding scope.

Use Debouncing#

For file-save triggers in active projects, 30-60 second debounces prevent trigger storms. A single focused coding session should not spawn dozens of runs.

Write Clear Instructions#

Think of instructions as guidance for a capable junior developer. Be explicit about:

• What the agent should and should not change
• How to verify success
• When to stop and flag for human review

Pin Cursor Version#

In team environments, pin your Cursor version for stability. The automations API is still in beta.

Comparing to Claude Code#

Both Cursor Automations and Claude Code offer agentic coding, but the execution model differs.

The practical split: Automations for background maintenance (dependency updates, PR prep, error fixing). Claude Code for foreground work (feature building, complex refactors, exploratory coding).

Many teams use both.

Getting Started Checklist#

1. Create .cursor/automations/ directory in your project
2. Start with one automation (test failure or dependency audit)
3. Set debounce_ms appropriately for your trigger type
4. Write instructions as guidance for a capable junior developer
5. Run with auto-apply disabled
6. Review staged diffs for 2-4 weeks
7. Enable auto-apply for trusted, narrow automations
8. Expand scope gradually

Official Sources#

FAQ#

What is Cursor Automations?#

Cursor Automations is a feature that lets AI coding agents run in the background based on triggers like test failures, file saves, PR opens, cron schedules, or webhook calls. Instead of waiting for manual prompts, automations execute autonomously and stage results for your review.

How much does Cursor Automations cost?#

Automations are included in Cursor Pro ($20/month), Pro+ ($60/month), Ultra ($200/month), and Business ($40/seat/month) plans. Usage is subject to monthly run quotas depending on your tier.

Can Cursor Automations run without my approval?#

By default, no. Changes are staged as diffs that require your review before applying. You can enable auto-apply for specific automations, but this is not recommended until you have built trust in the agent's judgment over several weeks.

How do Cursor Automations compare to GitHub Actions?#

Automations complement CI/CD rather than replacing it. GitHub Actions runs deterministic scripts. Automations run AI agents that can reason about code, propose fixes, and adapt to context. The common pattern is triggering automations from GitHub Actions for AI-assisted analysis on PRs.

Can I use Cursor Automations with Claude Code?#

Yes. Automations handle background maintenance inside Cursor (dependency updates, test fixes, PR prep). Claude Code handles foreground interactive work in the terminal (feature building, complex refactors). They serve different parts of the workflow.

What happens if an automation fails?#

Failed runs are logged with error output. The automation does not retry automatically. You can review the failure, adjust instructions, and re-run manually. Setting up monitoring via webhook triggers helps catch systematic failures.

How do I limit which files automations can modify?#

Instructions should explicitly scope what the agent can and cannot touch. For additional safety, use sandbox configurations that limit write access to specific directories or file patterns.

Are automations available on the free tier?#

No. Automations require a paid Cursor subscription (Pro or higher). The free Hobby tier includes basic AI features but not background automations.

Continue Reading#

• 303 AI Skills for 12 Careers: The Free Directory
• Buzz: Block's Agent-Native Messaging Layer on Nostr

Sources#

• Cursor Automations Launch Announcement
• Cursor Automations Guide
• Cursor Documentation
• Cursor Pricing