GitHub has added Anthropic’s Claude and OpenAI‘s Codex as coding agents you can use directly inside the platform. Unlike implementations where chat interfaces are bolted onto the side, these agents work where your code actually lives: in repositories, issues, and pull requests.
The feature is rolling out in public preview for Copilot Pro Plus and Copilot Enterprise subscribers. You can access it from GitHub’s web interface, the mobile app, and Visual Studio Code. No extra subscriptions are needed, though each agent session counts as a premium request.
GitHub already lets developers switch between different AI models when using Copilot, but this goes further. You’re not just picking a model to answer questions or generate snippets. You’re assigning actual work to agents that operate independently and produce code that goes through your normal review process.
The platform calls this Agent HQ. You can hand an issue to Copilot, Claude, or Codex and let them work on it. Each agent can submit draft pull requests. You can even assign the same issue to multiple agents simultaneously and compare their approaches side by side.
This setup treats agents as discrete workers rather than infinite conversation partners. Each session has a scope, produces specific artifacts, and leaves a trail. Everything stays tied to your repository instead of scattered across different tools.
Why GitHub Is Using Multiple Agents
Different AI systems handle different tasks better. One might excel at refactoring legacy code while another is stronger at writing new features from scratch. Rather than forcing teams to pick a single AI provider for everything, GitHub is letting developers choose the right tool for each job.
Mario Rodriguez, GitHub’s chief product officer, says the goal is reducing context switching. When you jump between your IDE, a chat interface, and your repository, you lose momentum. Keeping everything in one place means your work history, conversation threads, and code changes all live together.
Microsoft owns GitHub and has poured money into OpenAI, so opening the platform to competing agents is a unique choice. It appears GitHub cares more about being the central hub for development work than about pushing any particular AI system.
How It Works
Agents operate asynchronously by default. You start a session, assign a task, and the agent runs in the background. You can watch in real time or check back later. When it’s done, you get concrete outputs: code changes, comments on pull requests, or proposed solutions.
You can also bring agents into existing pull requests by mentioning them in review comments. Type @claude or @codex, describe what you want checked, and the agent responds with its analysis. This creates a transparent record of what the AI suggested and why.
The Claude integration emphasizes reasoning and iteration. Katelyn Lesse from Anthropic says the focus is on helping teams move faster while maintaining confidence in their decisions. Claude can explain its reasoning for proposed changes rather than just dumping code.
Codex has an interesting history here. The original Codex model helped build GitHub Copilot in the first place. Now it’s back as a standalone agent you can directly compare against Copilot, which has since evolved beyond its Codex origins.
How Will This Affect Development Teams?
The immediate benefit is comparison. You can see how different agents tackle the same problem and choose the approach that makes the most sense. Over time, teams will learn which agents handle which tasks best and develop preferences based on actual performance rather than marketing claims.
There’s also a governance angle. When AI-generated code lives outside your repository, tracking what was created and why becomes messy. Agent HQ keeps everything visible in the same place you review human contributions. The same standards, the same process.
GitHub plans to expand access beyond Pro Plus and Enterprise tiers eventually. They’re also working with Google, Cognition, and xAI to add more agents to the platform. GitHub really wants to be where AI agents do development work, regardless of who makes them.
Interestingly, Microsoft has been testing Anthropic’s Claude Code tool internally, asking developers to compare it with Copilot. The feedback apparently goes toward improving Copilot itself. So even as GitHub opens up to competition, the insights flow back into their own tools.
