A RESEARCH-BASED STARTING PATH
Choose AI help for an existing project
Choose the job first: changing code, reviewing a change or testing a user journey. These are complementary tasks. An agent’s completed change is a starting point for review, not automatic release approval.
Use a coding agent to investigate or change code
Codex, Claude Code, Cursor and GitHub Copilot support different repository and editor workflows. Check your languages, client, plan and organization policies.
Give the agent a specific task and the project’s own development instructions. Keep the change small enough to inspect. Run the relevant development checks before merging it.
Use review tools for an additional review pass
CodeRabbit can add contextual review feedback. GitHub Copilot also has review workflows whose availability depends on the plan and client. Evaluate suggestions alongside human review and tests.
Check what repository access is required. An AI review cannot guarantee that a change is correct or secure.
Use browser tests to check repeatable journeys
Playwright automates web checks across supported browser engines. It is useful when you can define correct behavior and maintain the test code and data.
Choose important journeys with meaningful assertions. Keep test state isolated. A test that passes is evidence about its assertions and inputs. It does not establish complete product coverage.
Keep a durable change history
Use version control and inspect the diff. Cursor’s checkpoint guidance distinguishes agent checkpoints from permanent Git history. Keep a recovery point before broad changes.
Review permissions, commands and external integrations before enabling them. Match the amount of autonomy to how well you can verify the result.