GitHub Copilot can now approve pull requests, a feature currently available in public preview for GitHub Copilot Pro, Pro+, Max, Business, and Enterprise plans. ^[1] On 2026-09-02, GitHub announced a series of updates to its AI coding assistant GitHub Copilot aimed at reducing token usage and operational costs without compromising task quality, as detailed in a blog post by GitHub engineers Erik Kristensen and Napalys Klicius. ^[2] Loop engineering is the discipline of designing workflows where AI agents run unattended to build, test, and fix software, replacing the need for manual code writing. ^[3] An approval assessment alone does not count toward merge requirements unless the feature is explicitly enabled by administrators. ^[4] Every Copilot code review now includes an approval assessment in the overview comment to signal whether Copilot considers the pull request ready to approve. ^[5] When enabled, Copilot can submit an approval that counts toward the repository's required-approvals rule, but its approval is dismissed if new commits are pushed after the approval. ^[6] GitHub engineer Erik Kristensen stated on 2026-09-02 that the goal of Copilot's efficiency update is 'fewer tokens per completed task' rather than fewer tokens per call, shifting focus from individual token use to end-to-end task efficiency. ^[7] GitHub engineers reported that batching results from background tasks such as long-running shell commands reduced token-related usage by 2.3%. ^[8] GitHub engineers reported that shortening Copilot's internal prompts without affecting critical model behaviors resulted in a 2.9% cost reduction per active hour in controlled experiments. ^[9] AI coding assistants like Claude Code, Cursor, and GitHub Copilot have moved past single-shot prompting to support autonomous agents that work through backlogs unattended. ^[10] The bottleneck in AI software development has shifted from the ability of AI to write code to the ability of humans to trust AI to run unwatched. ^[11] Yash Thakker, founder and CEO of AISOLO Technologies, stated that loop engineering involves setting up programs that call AI agents on a schedule or trigger, such as every 15 minutes or when a build fails. ^[12]