AI and software development: the describe, review, merge workflow

AI and software development converge most usefully in a three-step rhythm: describe, review, merge. You describe a task in plain language, an agent turns it into a pull request, and you review and merge it. The workflow keeps humans in control at the one point that matters, the merge, while delegating the mechanical middle to the agent. Tests and CI act as the automated gate that catches mistakes before a human ever looks.

This is a meaningful change from how AI was used a year or two ago, when the model produced suggestions you manually stitched into the codebase. The new shape moves the human from author to editor.

What changes and what stays the same

What changes is who does the typing. The agent reads the repository, plans the change, edits multiple files, and runs the suite. What stays the same is the discipline around it: branch protection, required checks, code review, and a deliberate merge. The agent slots into that existing structure rather than replacing it.

• Describe. Write the task the way you would write a ticket. Attach an issue or a failing test if you have one. Clarity here is the highest-leverage thing you do.
• Review. The agent opens a PR with the plan, the diff, and the test results. You read the plan first, then the diff, then the tests.
• Merge. You merge when you are satisfied. Nothing reaches main without this step.

The full loop, including the plan and test stages between describe and review, is laid out on how it works.

Tests and CI as the gate

The reason this workflow scales is that the agent runs your tests before you see the change. A pull request that fails CI is not handed to you as finished work; the agent reads the failure and revises. That makes your test suite the most important asset in an AI workflow, because it is the automated reviewer that runs on every change.

The better your tests, the more you can safely delegate. CI is not an afterthought in an AI workflow; it is the gate that makes delegation safe.

This is why teams that invest in automated testing get more out of an agent. Coverage turns into leverage: each well-tested module is a place the agent can work with confidence and you can review quickly.

Keeping humans in control

Control is not a feeling, it is a set of concrete checkpoints. Agentcode never merges on its own, never pushes directly to a protected branch, and never trains on your code. Every change is a reviewable PR on your existing GitHub or GitLab repository. The plan is visible before any edits, so you can redirect the agent early instead of discovering a wrong turn at review time.

This is the difference between automation that replaces judgment and automation that scales it. The agent removes the repetitive steps; you keep the decisions. For a step-by-step approach to the review itself, see how to review an AI pull request.

How the workflow plays out for a team

Consider a typical week. A backlog has a dozen small items: a flaky test, a dependency bump, a missing validation, a rename that touches twenty files. Historically these compete with feature work for senior attention and lose. In a describe-review-merge workflow, each becomes a described task. The agent opens PRs overnight, CI runs, and the team reviews a batch in the morning. The boring work moves without anyone context-switching into it.

1. An engineer files the backlog items as tasks for the agent.
2. The agent plans, edits, and runs tests for each, opening one PR per task.
3. CI gates every PR; failures are revised before review.
4. The team reviews the batch and merges the ones that pass the bar.

The result is more shipped pull requests at the same review standard. See how this looks for different audiences under engineering teams and solo developers.

Adopting it without disruption

Because the agent works on your existing repository and CI, adoption does not require a migration or a new place for code to live. You start by giving it one well-scoped task, review the PR exactly as you would review a colleague's, and expand from there. The workflow is additive: nothing about your current review process has to change. For the broader picture of where this sits among other tools, read the best AI coding tools guide, and for the foundational concept, what an AI coding agent is. Pricing for getting started is on pricing.