In a post published on the GitHub Blog, the company describes how integrating artificial intelligence agents can assist software developers across four central stages of the development process: Plan, Build, Verify, and Observe. According to data presented by GitHub, during October the platform recorded more than 1.2 billion commits in a single week—representing an increase of more than 400% compared to the same period a year earlier, and more commits than GitHub received in all of 2025. However, the post emphasizes that while accelerating implementation with AI increases the volume of generated code, it introduces pressure at other points in the process. In this context, a study by Faros AI is cited, which found that pull requests spent 441% longer in the review stage when development teams shifted from low AI adoption to high AI adoption.
Plan: Centralizing Context and Collaborative Decision-Making
According to GitHub, the planning process for an application programming interface (API) change often involves scattered details: a design document may live in one tool, a compatibility requirement is buried in Slack, and an old pull request explains why the existing implementation was built the way it was. Searching for these details, comparing approaches, and figuring out how existing clients will keep working can consume hours of work before implementation even begins.
To address this challenge, GitHub suggests using Copilot within the communication tools where the team works, such as Slack or Microsoft Teams, to collaborate with product managers, designers, analysts, and business stakeholders. The tool allows teams to pull context from connected tools, check underlying assumptions, contribute ideas, and refine requirements together.
Once requirements are agreed upon, Copilot can be prompted to create GitHub Issues that include the context and acceptance criteria needed for implementation, breaking the work down into tasks that can be handled separately. In addition, GitHub Agentic Workflows enable labeling issues and estimating task complexity. Low-complexity tasks can be assigned directly to Copilot, allowing developers to dedicate their time to more complex decisions. For complex work, Plan Mode in a preferred Copilot interface enables comparing different approaches and refining an implementation plan, with the developer able to review and approve the plan before handing it off to Copilot for implementation.
Build: Parallel Progress and Maintaining Cross-Device Continuity
In the implementation phase of turning a plan into code, developers often encounter interruptions, meetings, and edge-case requirements that make single-focus work difficult. GitHub presents mechanisms within Copilot designed to allow work to progress in parallel:
First, session continuity allows developers to continue the same Copilot conversation across different clients and devices. Copilot runs tasks in cloud-based sandboxes, keeping execution off the developer's local machine and allowing them to leave or close their laptop and return to the session later.
Second, the Copilot application gives each session isolated copies of the repository, ensuring an implementation task and a documentation update do not share an unfinished working tree.
In addition, project knowledge is retained and carried into each task: repository instructions document conventions and validation commands, while shared memory stores useful facts that Copilot can reuse across the repository. A tool named Copilot Dreaming reviews eligible sessions and pull request feedback to keep memories current and propose updates to instructions or skills.
Regarding model selection, Copilot Fusion allows developers to select models and reasoning depth for the task, and even bring in additional models to review a solution when needed. The developer sets the effort level to balance quality, cost, and response time. Furthermore, enterprise managed settings enforce company policies and configured tools across all Copilot surfaces.
Verify: Managing Pull Requests and Code Quality
The strain created by accelerated code generation is felt when reviewing large, complex pull requests. GitHub notes that review wait times can be shortened in two main ways: making changes easier for human review, and using Copilot to clear routine checks and fixes before a human reviewer opens the diff.
The suggested approach includes using stacked pull requests instead of creating a single massive change. Breaking dependent work into smaller increments allows each pull request to be reviewed on its own, maintains visibility into dependencies, and enables other teams to build on top of the work before it merges.
Alongside this, GitHub Code Quality checks changes against coding practices and organization standards before they reach a human reviewer. Copilot code review provides automated feedback, allowing users to select lighter analysis for simple changes or deeper analysis for complex changes, either manually or via automatic Copilot selection.
In security, AI Scan locates vulnerabilities in languages and files beyond the coverage of CodeQL, and AI secret detection helps identify unstructured secrets such as passwords. Furthermore, license compliance checks dependency licenses against company rules prior to merging, and the GitHub package firewall, currently in public preview for npm, can block downloads that violate configured security policies.
In the remediation stage, Agentic Autofix proposes fixes for security issues, while Agent Merge automatically handles failing checks, review comments, and merge conflicts. If configured by the user, Copilot can even merge pull requests on its own once verification steps have passed.
Observe: Learning from What Happens After Production Deployment
The fourth stage, according to the post, addresses what happens after code is deployed. Situations where intermittent timeouts appear and customers report issues require data-driven investigation.
GitHub presents the ability to connect operational evidence with engineering context: with access to observability tools and business data through connectors such as Fabric, Copilot can assist in investigating which requests or customer groups are affected by an issue.
Developers can then ask Copilot to draft an issue accompanied by supporting evidence and acceptance criteria, including a regression test for the uncovered failure. In addition, Dreaming allows teams to extract lessons for future work by using sessions and repository activity to refine memories, skills, and code review guidance.
How to Get Started with Agentic Work
The post recommends that developers start with a single task already on their backlog and evaluate it beyond the coding activity alone—examining what context it requires, how it will be tested and reviewed, and how they will know it functioned properly after deployment. The recommendation is to start with whichever item causes the most frustration that week, apply an agent to it, and gradually expand adoption to other parts of the development workflow. This topic was also presented at GitHub Universe 2026 in the session "Agentic Engineering: Past the Pilot."