safe-agentic-workflow / rules

---
description: "Guidelines for Cursor background agents running long tasks autonomously in isolated VMs."
alwaysApply: false
---

# Background Agent Guidelines

Cursor Background Agents run long tasks autonomously in isolated Ubuntu VMs with internet access. They clone the repo from GitHub, work on a separate branch, and open PRs when complete.

## How Background Agents Work

- Each background agent runs in an isolated Ubuntu VM provisioned by Cursor
- The agent clones the repository from GitHub (it does NOT use your local checkout)
- It works on a dedicated branch and can commit, push, and open pull requests
- You are notified when the task completes or if the agent needs input
- Multiple background agents can run in parallel on different tasks

## SAFe Gate Chain Compliance

Background agents MUST follow the same SAFe workflow as interactive sessions:

1. **Stop-the-Line Gate**: Verify AC/DoD exists before starting work
2. **Pattern Discovery**: Search `patterns_library/` before implementing
3. **Commit Format**: Use SAFe format `type(scope): description [{{TICKET_PREFIX}}-XXX]`
4. **Branch Naming**: Use `{{TICKET_PREFIX}}-{number}-{description}`
5. **Exit States**: Declare the correct exit state when complete

## PR Creation from Background Agents

When a background agent opens a PR:

- Reference the Linear ticket in the PR title: `feat(scope): description [{{TICKET_PREFIX}}-XXX]`
- Follow the PR template from `CONTRIBUTING.md`
- Include a test plan in the PR body
- Tag appropriate reviewers per the 3-stage review process
- The PR still requires QAS validation and HITL merge authority

## Best Practices

- Assign one ticket per background agent for clear scope
- Provide the full spec path: `specs/{{TICKET_PREFIX}}-XXX-feature-spec.md`
- Reference the agent role: "Act as the BE Developer" or "Act as the FE Developer"
- Include validation commands the agent should run before opening the PR
- Monitor agent progress via Cursor's background agent dashboard

## Limitations

- Background agents cannot access your local filesystem or environment variables
- They start fresh from the GitHub remote each time
- They cannot interact with local Docker containers or databases
- Session state does not persist between background agent runs
- They inherit repository-level Cursor rules but not user-level settings

## When to Use Background Agents

| Scenario | Use Background Agent? |
|----------|-----------------------|
| Implementing a well-specified story with clear AC | Yes |
| Exploratory debugging or investigation | No -- use interactive |
| Multi-file refactoring with clear scope | Yes |
| Tasks requiring local environment (Docker, DB) | No -- use interactive |
| Running test suites and fixing failures | Yes |
| Architectural decisions requiring discussion | No -- use interactive |

## Example Invocation

```
Run as a background agent:
- Act as the BE Developer (see .claude/agents/be-developer.md)
- Implement {{TICKET_PREFIX}}-42 per specs/{{TICKET_PREFIX}}-42-user-api-spec.md
- Follow patterns from patterns_library/api/user-context-api.md
- Run: pytest tests/integration/ && ruff check . && mypy .
- Open a PR when all checks pass
```

More agent context in bybren-llc/safe-agentic-workflow

60 other files this repository gives its agents.