GitHub Copilot Evolves into Extensible Platform
GitHub has significantly expanded its Copilot AI assistant, introducing an SDK for building custom agent skills and support for persistent, context-aware "agentic memory." The platform is also previewing support for Claude and Codex coding agents, signaling a more open model ecosystem. Mobile users can now also select their preferred AI model, offering more flexibility for different coding tasks.
- The new GitHub Copilot SDK, released in technical preview, allows developers to build agent-style workflows into their own applications using the same engine that powers the Copilot CLI. It initially supports Node.js, Python, Go, and .NET. - "Agentic memory" addresses the problem of "contextual amnesia" by allowing Copilot to learn and retain repository-specific coding patterns and conventions. These "memories" are stored per-repository, can be reviewed or deleted by repository owners, and automatically expire after 28 days to avoid using stale information. - The preview of Anthropic's Claude and OpenAI's Codex agents is available to users on Copilot Pro+ and Copilot Enterprise plans. This allows developers to assign a task to multiple agents simultaneously and compare how each one approaches the problem. - During the public preview period, each session initiated with the Claude or Codex agents consumes one "premium request" from the user's subscription. - The model picker on GitHub Mobile is for Copilot Pro and Pro+ subscribers and includes options like Anthropic Claude Opus 4.5/4.6, Claude Sonnet 4.5, and OpenAI GPT-5.1/5.2-Codex, in addition to an automatic selection mode. - Developers can create reusable "Agent Skills" by creating folders with specific instructions and resources. Copilot will then automatically detect and apply these skills when they are relevant to a given task, enabling more specialized and repeatable assistance.
