Reliable Software Workflows

A starting question

An AI assistant produces a feature that works on your laptop. What has to happen before it is safe for a team to ship?

The change must be understandable, checked, deployed deliberately, and observable afterward. AI can speed up writing, but it does not remove the need for a delivery workflow.

Keep changes reviewable

A focused diff is easier to reason about than a feature mixed with unrelated refactoring. Explain what changed and why. Ask another developer to examine the code and important edge cases when team practice calls for it. A pull request provides a place for the diff, discussion, and checks; GitHub’s documentation describes those parts.A reviewer should still understand the behavior rather than approve because the author used AI.¹

Run checks before release

Automated tests exercise specified behavior. Continuous integration, or CI, runs checks such as tests and builds when code changes. Continuous delivery/deployment, often shortened to CD, prepares or publishes a release through automation; teams may still require human approval. GitHub Actions is one example of a CI/CD system.The workflow is a practice, not a requirement to use a particular platform.²

For an AI-assisted cart change, CI might run a cart-total unit test, an API integration test, and an end-to-end add-to-cart flow. Those checks reduce uncertainty, but they cannot cover inputs nobody anticipated.

Watch what happens after deployment

Monitoring means observing whether the running system behaves as expected. Logs record events, metrics summarize values such as error rates, and traces can follow a request through multiple components. OpenTelemetry’s primer explains these signals.Choose signals tied to user-visible behavior, not just “the server is running.”³

If the cart fails after release, rollback means returning to a known working version or disabling the risky change through an established process. A rollback is not an admission that review failed; it is a recovery option. Teams need to know what was released and how to reverse it before an incident, not during one.

Recap and try it

Reliable delivery turns a plausible patch into a change people can inspect, test, observe, and recover from. Apply the same bar to human- and AI-written code.