Ing. Calogero Bono
GitHub has published a developer guide to agentic engineering, meaning working with AI agents across the entire software lifecycle. The starting point is the numbers the company reports. During one week in October, commits exceeded 1.2 billion, more than 400% higher than a year earlier. Writing code now costs less, but planning, reviewing and observing remain the bottleneck, and that is where real productivity is won or lost.
A few terms before we get into the facts. Agentic engineering is a way of working in which AI agents carry out tasks autonomously while you steer them. A pull request is a proposed change that a colleague reviews before it is merged. Triage is the sorting of reports by priority. CodeQL is GitHub's tool that analyzes code for vulnerabilities. A sandbox is an isolated environment where code runs without touching your computer.
Why doesn't more code mean more productivity?
According to GitHub, one week in October saw more commits than in all of 2025. The company asks whether developers really feel 400% more productive.
The second figure comes from a Faros AI study cited by GitHub. Pull requests stayed in review 441% longer when teams moved from low to high AI use. The work shifts from writing to verifying.
It is the same scenario we covered when discussing GitHub Rubicon, designed for many agents on a single repository.
How do you plan with Copilot?
The Plan phase is meant to gather in one place the context scattered across tools, chats and old pull requests. GitHub points to Copilot in Slack or Teams for planning together with product managers, designers and analysts.
- Copilot can create GitHub Issues with acceptance criteria and separable tasks.
- GitHub Agentic Workflows labels issues and estimates complexity, so simple tasks go straight to Copilot.
- Plan Mode compares approaches and produces a plan to approve before implementation.
How do you work on several tasks at once?
In the Build phase, session continuity lets you resume the same Copilot session from different devices. Tasks run in cloud sandboxes, so they keep going with your laptop closed. The Copilot app gives each session an isolated copy of the repository.
Project knowledge flows through repository instructions and shared memory. Copilot Dreaming reviews eligible sessions and pull request feedback to update the memories. With Copilot Fusion you choose models and reasoning depth, balancing quality, cost and time. An approach close to the one described in agentic AI and advanced automation.
How do you cut pull request review time?
For GitHub, the review queue is "441% in practice". The answer is to work on two fronts, changes that are easier to review and routine checks handed to Copilot.
- Stacked pull requests split dependent work into small pieces.
- GitHub Code Quality checks changes against organization standards.
- Copilot code review adjusts its depth to the size of the change.
- AI Scan looks for vulnerabilities beyond CodeQL's coverage, while AI secret detection finds unstructured secrets such as passwords. For more, see our guide to secrets management.
- The package firewall, in public preview for npm, can block downloads that violate policies.
- Agentic Autofix proposes security fixes and Agent Merge handles failed checks, comments and conflicts.
If you configure it, Copilot can also merge pull requests once the checks pass.
Limits to know
The text is a guide published by GitHub, so the benefits are the company's own claims and not independently measured results. In the material available to us, the Observe phase is cut off, so we cannot describe it in full. Pricing, required plans and the availability of each feature are missing, apart from the public preview of the package firewall. The Faros AI study is cited, not verified by us.
What changes for those working in digital
If you develop with AI, the time you save writing code risks being lost in reviews. It makes sense to work with small, focused pull requests and to automate routine checks before a colleague gets involved. Anyone using Cursor or other assistants should apply the same reasoning. For small teams, writing good repository instructions becomes part of the job.
Our take
The guide deserves credit for starting from a real problem, review, instead of boasting only about speed. It is still promotional material, however. Many of the features mentioned, from Agent Merge to automatic merging, shift trust toward the agent precisely where more human attention is needed. The risk is replacing the review queue with a queue of automated checks that nobody reads. The method has merit, but it should be adopted one step at a time and measured.
Frequently asked questions
What is agentic engineering according to GitHub?
It is a workflow in which AI agents help across four phases, Plan, Build, Verify and Observe. The developer guides the outcomes. GitHub presents it as an extension beyond code generation alone.
How many commits did GitHub record?
More than 1.2 billion in a week in October. According to GitHub, that is over 400% more than a year earlier and more than all of 2025 combined.
What does the Faros AI study say?
Pull requests stay in review 441% longer with high AI adoption compared with low adoption. GitHub cites it as evidence that the pressure is shifting to review.
Can Copilot merge pull requests on its own?
Yes, if you configure it to. GitHub says Copilot merges pull requests once the checks you have set up pass.
Is the package firewall available to everyone?
It is in public preview for npm. It can block downloads that violate the security policies you configure. For other ecosystems the source says nothing.
What to watch in the coming months for assisted reviews
The question is whether these tools actually reduce the wait for reviews. Independent data will tell us more than guides will.
- Measure the average review time of your pull requests before and after introducing agents.
- Check when the Observe phase will be described in full and how it connects to your monitoring tools.
- Verify the plans and costs of the security features before relying on them for your checks.
