Beyond code generation: A developer’s guide to agentic engineering

GitHub's guide to agentic engineering opens on a gap between output and progress. GitHub says it saw more than 1.2 billion commits in one week this October, over 400% more than a year earlier, while a Faros AI study it cites found that pull requests spent 441% longer in review when teams went from low to high AI adoption. Its argument is that faster code generation moves the pressure onto the rest of a developer's job, so it splits that job into four stages, planning, building, verifying and observing production, and assigns agents and Copilot features to each.

Verification is where the argument is most concrete, because GitHub presents the swollen review queue as that 441% figure in practice. It proposes working on both ends of the queue. Authors and their agents can use stacked pull requests to break dependent work into smaller changes that are each reviewed on their own, instead of one change touching dozens of files. Automated tools then clear routine problems before a person opens the diff: GitHub Code Quality checks against the organization's standards, Copilot code review scales its analysis to how complex the change is, and Agent Merge addresses failing checks, review comments and merge conflicts. If configured, Copilot will merge a pull request once the verification steps pass. The intended result is that human reviewers spend their attention on whether the code does what was intended.

The other stages follow the same pattern of moving routine work to agents. In planning, Copilot can turn agreed requirements into GitHub Issues with acceptance criteria, and GitHub Agentic Workflows can label issues and estimate complexity so developers can assign low-complexity tasks directly to Copilot. In building, Copilot can run tasks in cloud sandboxes, the Copilot app gives each session its own isolated copy of the repository, and a feature called Copilot Dreaming reviews past sessions and pull request feedback to keep shared memories current and propose updates to instructions. After release, connectors to observability tools let Copilot help work out which requests or customers a problem affects and draft an issue with a regression test. Two of the products the guide relies on, the Copilot app and Copilot Fusion, which picks models for a task, were among those GitHub presented at Universe 2026.

This is GitHub describing its own products, and the post offers no measurement that they shorten review or raise productivity. The 441% figure describes the problem as Faros AI measured it, not the effect of GitHub's fixes. GitHub itself doubts that commit volume equals productivity, asking whether developers feel 400% more productive than last year, and its practical advice is modest: pick the part of the process that frustrates you most this week, put an agent on it, and expand from there.