Microsoft Agent Framework, the open source orchestration library for building multi-agent AI systems, now ships generally available connectors for GitHub Copilot and Claude Code. The connectors let .NET and Python agents delegate coding tasks to either system without writing custom adapter code.
The project has grown to 12,300 stars on GitHub with 2,100 forks and over 2,600 commits, putting it in the same league as other major agent frameworks. The connectors add a capability that the existing frameworks largely leave as an exercise for the developer.
The update is significant because it turns the Agent Framework from a framework for orchestrating AI agents into a framework that can also use AI coding tools as execution engines. Your orchestrator agent can decide a task needs code written, hand it to GitHub Copilot or Claude Code via a standard connector, and get the result back without managing subprocesses or API integration yourself.
What the connectors actually do
Before this release, building an agent that could use GitHub Copilot or Claude Code meant writing your own adapter. You had to handle authentication, session management, prompt formatting, result parsing, and error handling. Every team that wanted this capability built it slightly differently, which meant maintenance burden and inconsistent behavior.
The Agent Framework connectors standardize the interface. You define a coding task, the connector routes it to the configured tool, and the result comes back as a structured response that your agent can process. The framework handles the boilerplate.
The connectors plug into the Agent Framework’s built in code execution model. That means they are not standalone wrappers. They work within the same middleware pipeline, session management, and observability that the framework provides for other agent types. If you already have monitoring set up for your agents, the coding connectors report through the same channels.
How this compares to other agent frameworks
LangChain and CrewAI both support code execution, but they approach it differently. LangChain has a Python REPL tool that runs arbitrary code in a sandboxed environment. CrewAI delegates to subagents that can write and execute code. Neither has a built-in connector for a specific AI coding assistant like GitHub Copilot or Claude Code. You would need to build that integration yourself.
Semantic Kernel, Microsoft’s other agent framework, has code execution capabilities through its kernel functions and plugins, but it does not have dedicated connectors for GitHub Copilot or Claude Code either. The Agent Framework connectors are the first Microsoft offering that treats AI coding assistants as first-class execution targets within an agent orchestration pipeline.
The difference matters for teams that want to use AI coding tools as production components, not just developer productivity aids. If your agent needs to generate code as part of its output, having a standardized connector means you can swap the coding backend without changing your agent logic. The interface stays the same whether Copilot or Claude Code handles the task.
Why this matters for multi-agent workflows
The most interesting use case is delegating coding subtasks from a planner agent. Your orchestrator can decompose a complex request into research, coding, and verification steps, hand the coding step to a GitHub Copilot or Claude Code worker agent, and move on to coordinating the other subtasks. The coding happens in parallel with the other work.
This is different from the typical “single developer with an IDE” Copilot interaction. The agent is not suggesting code for a human to accept. It is generating code to be integrated into a larger automated pipeline. The code might be a batch transformation script, a test suite, a configuration file, or a small microservice that gets deployed as part of the agent’s output.
The connector model also means you are not locked into one coding AI. You can use GitHub Copilot for some tasks and Claude Code for others, or route tasks based on the programming language, the complexity, or the expected output format. The agent decides which tool to call based on its instructions.
Practical considerations
The connectors are generally available, so they are past the preview stage. That means API stability, support commitments, and production readiness. But GA does not mean every edge case is handled. The coding connectors work best for well defined, self contained coding tasks. If your agent needs to refactor an existing codebase across multiple files, that is harder than generating a new file from a specification.
Authentication flows through the existing Agent Framework credential management. If you already have GitHub Copilot or Claude Code configured for your project, the connector should pick up the existing credentials. No separate setup step for the connector itself.
Cost is worth considering. Each coding task consumes a GitHub Copilot or Claude Code API call on top of the LLM calls your orchestrator agent makes. If your agent pipeline generates code for every user request, the API costs add up faster than if you rely on the agent’s own code generation capability. Use the connector when you need specialized coding output, not for every task.
Getting started
The connectors ship as part of the Agent Framework SDK. Install or update to the latest version of the Python or .NET package and look for the GitHubCopilotConnector or ClaudeCodeConnector classes. The framework documentation has examples for setting up a coding worker agent and integrating it into an existing orchestration pipeline.
Start with a narrow use case. Generate a single file from a specification, verify the output, and inspect the code before giving the agent more freedom. The connectors handle the integration, but AI generated code still needs human review when it goes into production.