GitHub Universe 2025 opening keynote

Key Concepts

• GitHub Copilot Coding Agent: An AI-powered coding assistant integrated into the developer workflow, extending beyond simple autocompletion to execute and coordinate tasks.
• Agent HQ: A centralized platform on GitHub designed for developers and AI agents to collaborate, manage, and steer tasks. It aims to be the "home" for all coding agents.
• Mission Control: A feature within Agent HQ providing a single view to monitor, manage, and review all active agent sessions across projects.
• Custom Agents: AI agents that can be tailored for specific scenarios, such as writing documentation or performing test-driven development.
• Model Context Protocol (MCP): A protocol that allows AI models to understand the entire codebase and development environment, not just individual files.
• GitHub Copilot Code Review: An AI-powered code review system that provides meaningful reviews of pull requests, identifying bugs and security issues.
• Plan Mode: A feature in VS Code that allows developers to collaborate with Copilot to create step-by-step plans for tasks before writing code.
• AI Controls (Enterprise): Tools and features for enterprise administrators to manage Copilot deployments, set security policies, audit agent activity, and control agent access to codebases.
• Agentic Programming: A paradigm where AI agents are integral collaborators in the software development process, working alongside human developers.

GitHub Universe 2023: The Agentic Era of Software Development

This summary details the key announcements and themes from GitHub Universe, focusing on the integration of AI agents into the developer workflow, transforming how software is built. The core message revolves around GitHub becoming the central platform for both developers and AI agents, fostering a new era of collaborative development.

1. The Evolution of Developer Pain and GitHub's Role

Kyle Daigle opened by reflecting on his early days at GitHub, highlighting the inherent humility in software development. He emphasized GitHub's mission to solve developer pain and tame chaos, noting the platform's growth to 180 million developers, with one joining every second. The presentation then addressed the current challenge: despite AI's promise to increase productivity, developers are often burdened by increased complexity and fragmented workflows across multiple tools. GitHub's strategy is to integrate AI natively into existing developer processes, allowing choice and flexibility.

2. GitHub Copilot as a Fully Fledged Coding Agent

The core of the presentation focused on the evolution of GitHub Copilot from an autocompletion tool to a powerful coding agent.

• Deeper Context and Reasoning: Copilot now operates with a deeper understanding of the codebase, enabling stronger reasoning capabilities.
• Task Execution and Coordination: It can execute and coordinate tasks across the entire developer workflow, moving beyond simple code generation.
• Integrated Workflow: Unlike "bolted-on" solutions, Copilot is deeply integrated into the existing GitHub Flow (issues, pull requests) and asynchronous workflows.
• Agent HQ and Mission Control: Developers can now manage multiple Copilot sessions in parallel through a new interface called "Agent HQ," with "Mission Control" providing a single view of all agent sessions, task progress, and code generation.
• Real-World Impact: The secret scanning team at GitHub used Copilot Coding Agent to build and ship validators for 90 new providers in one month, a task that would have taken months otherwise. Copilot is already the number one contributor to merged pull requests within GitHub's codebase.
• Control and Context: Copilot agents are designed to provide the right context about the developer and their work, with a strong emphasis on giving developers control.
• Extensibility and Customization:
  • MCP Integration: Copilot can connect to core services like Figma, Sentry, and Jira via their MCP Servers and the GitHub MCP Registry.
  • Self-Hosted Runners: Copilot Coding Agent can now run on self-hosted runners, providing flexibility for enterprise environments.
  • Integrated Security Tools: CodeQL, dependency scanning, and secret protection are integrated as tools Copilot can utilize.
  • Custom Agents: Developers can create custom agents tailored for specific scenarios like documentation writing or test-driven development.

Demo: Jessica Deen Showcases Copilot Coding Agent

Jessica Deen demonstrated the capabilities of the Copilot Coding Agent by building a companion website that generates custom wallpapers from GitHub profiles.

• Session Creation: Agent sessions can be initiated from GitHub issues, the CLI, or anywhere developers work.
• Mission Control: Used to manage multiple in-flight agent sessions, including the one for mobile wallpaper support.
• Custom Docs Expert Agent: Demonstrated the use of a specialized custom agent for writing documentation.
• Customizable Environments: Copilot's execution environment can be customized, allowing access to non-standard tools, internal dependencies, and cached dependencies.
• Mid-Session Steering: Developers can steer agents mid-session by providing specific instructions or dimensions.
• Internal Iteration and Verification: Copilot iterates internally, checking for code quality and security vulnerabilities using tools like CodeQL, Linters, and Secret Protection.
• Pull Request Integration: Completed agent sessions result in comprehensive pull requests with summaries, screenshots (via Playwright MCP), and diff views.
• IDE Integration: Sessions can be opened directly in VS Code or Codespaces, allowing seamless transition between web and editor environments.

3. Agent HQ: The Universal Platform for Coding Agents

The vision for Agent HQ extends beyond GitHub Copilot to encompass all coding agents.

• Open to All Agents: Agent HQ will welcome coding agents from various providers, including Claude, Codex, Google Cognition, and xAI.
• Unified Management: Provides a single place to start, steer, and manage sessions from all agents across the web, IDE, CLI, and mobile.
• Developer Control: Features like new branch controls ensure developers dictate when agent-generated code is run.
• Agent Identity: Each agent has a full GitHub identity, allowing for granular access control, assignment of issues/PRs, and mentions in comments.
• Enterprise-Grade AI Controls: Administrators gain tools to manage access, set security policies, and audit agent activity.

Partnership Demos:

• Anthropic's Claude: Mike Krieger showcased Claude as a native collaborator within Agent HQ.

  • Issue Assignment and Branch Creation: Claude can pick up issues, create branches, and submit pull requests.
  • Interactive Pull Requests: Developers can interact directly with Claude's pull requests, ask questions, and request changes.
  • Skills for Specialization: Claude can utilize "skills" (folders of prompts and scripts) to perform specific tasks, such as adhering to brand guidelines for marketing page updates.
  • Claude Agent SDK: Powers Claude's integration with core tools, context management, and permission frameworks.

• OpenAI's Codex: Alexander Embiricos demonstrated Codex integration within VS Code.

  • VS Code Insiders Preview: GitHub Copilot Pro+ users can access Codex in VS Code Insiders today.
  • Seamless Authentication: Sign-in to Codex is integrated with GitHub Copilot.
  • Parallel Agent Execution: Codex can run multiple agents in parallel for tasks like game development, adding features, and fixing bugs.
  • Optimized Model and Harness: Codex leverages GPT-5, optimized for agentic software engineering, and an open-source harness for prompt management, tools, and run loops.
  • Adaptive Reasoning Time: Codex can adapt its reasoning time based on task complexity, from quick answers to hours or days for highly complex problems, managed through "auto compaction" of the context window.

4. VS Code: The AI-Native Editor

The presentation highlighted the deep integration of AI features within VS Code, making it the premier AI-native code editor.

• Collaborative Development: GitHub, Copilot, and VS Code teams have been working together to build AI features in the open.
• Transparency and Trust: Prompts sent and data collected are inspectable, providing transparency for enterprise customers and the security community.
• Rapid Iteration: Weekly releases of VS Code and frequent Copilot improvements (nearly 200 in six months).
• Key Copilot Features in VS Code:
  • Full MCP Support: Understanding of the entire codebase.
  • Custom Rules: Developers define standards for Copilot.
  • Slash Commands: Sharing common prompts and workflows.
  • Enhanced Semantic Code Models: Deeper context and smarter code understanding.
• Model Choice and Auto Mode:
  • Fast Model Updates: New models like Claude Sonnet 4.5 and GPT-5 Codex are shipped as soon as they are released.
  • External Model Providers: VS Code extensions (Hugging Face, Azure AI Foundry) allow integration with other model providers (OpenRouter, LM Studio).
  • Auto Mode: Copilot automatically selects the best model for a given task.
• Plan Mode Demo (Pierce Boggan):
  • Collaborative Design Partner: Copilot analyzes the codebase and generates a step-by-step plan for implementing features, identifying gaps and missing decisions.
  • Test-Driven Development (TDD) Custom Agent: A custom agent was demonstrated that follows TDD principles, writing failing tests, implementing features, and ensuring tests pass.
  • MCP Integration in VS Code: VS Code is the only editor supporting the full Model Context Protocol specification, with tools from the GitHub MCP Registry (Stripe, Figma, Sentry) available.
  • Agent Sessions in VS Code: All agents are accessible within a single view in VS Code, allowing developers to kick off, monitor, and review sessions locally or remotely.
  • Asynchronous Task Execution: Developers can create asynchronous tasks for Copilot Coding Agent, which generates draft pull requests.
  • Applying Remote Changes Locally: Changes from remote agent tasks can be applied directly into the local codebase within VS Code.
  • Experimentation with Variations: Copilot can generate multiple variations of a feature asynchronously, allowing developers to experiment without tripling their workload.

5. Code Review: Enhancing Quality and Speed

The presentation emphasized the evolution of Copilot Code Review to address the challenges of reviewing large volumes of code.

• Meaningful Reviews: Copilot Code Review provides more context, tool calling (CodeQL, Linters), and finds more interesting bugs.
• Custom Instructions: Repositories can have tailored reviews based on custom instructions.
• Security and Deterministic Analysis: Combines model intelligence with CodeQL for deterministic analysis to catch security issues before merging.
• Applying Suggestions: Developers can apply suggestions individually or all at once.
• Automated Pull Request Creation: At-mentioning Copilot in a comment can trigger it to create a new pull request with the suggested changes, streamlining the process of fixing minor issues.
• Focus on Core Problems: This automation allows teams to focus on larger, more complex problems rather than chasing down minor fixes.
• Accessibility and Custom Banners: Demonstrated examples of Copilot fixing accessibility issues and adding custom banners to the wallpaper app.
• Enterprise Enablement: Agentic Copilot Code Review is available for enterprise admins to enable within preview features.
• Pull Request Experience Improvements: Ongoing work to make PRs faster, more reliable for large PRs, and easier to manage, including commenting on unchanged parts of files.

6. Enterprise Focus: Quality, Metrics, and Controls

A significant portion of the presentation was dedicated to addressing the needs of enterprise customers.

• Code Quality: GitHub Code Quality extends Copilot to assess the maintainability and reliability impact of code changes within pull requests, providing insights from the security lab and organizational summaries.
• Copilot Metrics Dashboard (Public Preview):
  • Impact Measurement: Shows the impact of using Copilot and any coding agent across the organization.
  • Guidance and Advice: Provides information on what to measure for driving real impact.
  • API Access: Allows for deep dives into data.
  • Availability: Available with existing Copilot Business or Copilot Enterprise subscriptions.
• AI Controls for Enterprise:
  • Centralized Management: A single enterprise-wide admin interface for managing Copilot deployments and future features.
  • Granular Control: Highly requested controls for managing agents, defining shared locations for custom agent profiles, viewing audit logs, and controlling agent access to codebases and features.
  • Custom Role: A new custom role for administrators and trusted delegates.

7. The New Era of Collaboration: Developers and Agents Working Together

The overarching theme is the emergence of a new era of collaboration between developers and AI agents.

• GitHub as the Platform: GitHub is positioned as the platform for this new era, unblocking developers to build what's next.
• Fireside Chat with Satya Nadella:
  • Evolution of Toolchains: Nadella discussed the historical shifts in developer toolchains, from assembly to compilers and now to AI agents.
  • Agent HQ's Importance: He emphasized the need for tools and patterns to effectively use AI agents, preventing chaos and enabling developers to "fall into the pit of success."
  • GitHub's Role: GitHub's primitives (Git, PRs, Issues, Actions) combined with new agentic capabilities will compose a powerful platform for agentic programming.
  • Microsoft's Developer Focus: Reaffirmed Microsoft's commitment to developer tools, platforms, and partnerships.
  • Skills for the New Generation: Nadella advised developers to focus on learning the "how" of using the new toolchain, emphasizing metacognition, continuous learning, and the emergence of new patterns and practices.

8. Conclusion: Building Together

The presentation concluded by reiterating GitHub's commitment to being the home for developers, regardless of their workflow. The message is clear: the integration of AI agents is not a future concept but a present reality, fundamentally changing how software is built. The call to action is for developers to embrace these new tools and build what's next, together.