Event in Spanish: Jueves de Quak con Carlos Alarcon y Copilot SDK

Key Concepts

• GitHub Copilot SDK: A toolkit for building custom AI agents, offering flexibility and control over model selection, tool integration, and cost optimization.
• Agent Creation: The process involves installing the SDK, establishing sessions with language models, defining and integrating tools, and utilizing system messages for agent personalization.
• Tools (MCPs & Skills): Custom functions that extend agent capabilities by interacting with external systems like databases and APIs, enhanced by “skills” for improved accuracy.
• Streaming & Error Handling: Techniques for improving user experience through real-time responses and robust error management.
• Cost Optimization: Leveraging BYOK (Bring Your Own Key) and selecting appropriate models (like Sonnet) to reduce expenses compared to pre-packaged solutions.

Introduction: Quicks & the Copilot SDK

The “Jueves de Cuac” broadcast features Carlos, founder of Quicks, discussing the new GitHub Copilot SDK and its application in creating AI agents. Quicks provides corporate training via instant messaging platforms (Slack, Teams, WhatsApp) for employees not consistently using laptops, enabling scalable and targeted learning for companies with over 1000 employees. Carlos notes Copilot has been a pioneer in code assistance, stating, “Copilot always has been very aware of all these proposals. Of the first code assistants in the world was Copilot.”

SDK Functionality & Architecture

The Copilot SDK is a set of libraries simplifying AI solution development using Copilot’s power. It’s built upon Copilot CLI, enabling interaction with language models via the command line. The SDK supports TypeScript, Python, Go, and .NET. The core architecture involves an application using the SDK client to connect to Copilot CLI, granting access to models from OpenAI, Google (Gemini), and others.

Key SDK Features

The SDK boasts several key features:

• Multi-model Support: Compatibility with models like GPT-4.1, Sonnet 4.5, and Gemini 3 Pro.
• Streaming Responses: Implementation of delta streaming for faster perceived response times.
• Session Management: Ability to manage multiple conversational sessions, each with unique models and tools.
• Tools Integration: Integration of custom functions (tools) written in Python (or other languages) to extend agent capabilities, enabling interaction with databases, APIs, and specific actions.
• Bring Your Own Key (BYOK): Option to use personal API keys (Azure, OpenAI, Cloud) instead of a Copilot subscription.
• Event Generation: The SDK generates events for tracking and reacting to interaction stages (user messages, AI responses, tool execution, errors).

Building Agents with the SDK: A Step-by-Step Process

Creating an agent with the SDK involves:

1. Installation: Using PIP or Poetry.
2. Client & Session Creation: Creating an SDK client and a session connected to a specific model (GPT-4.1 in the example).
3. Prompt Submission: Sending prompts using send_and_wait or send_and_collect (for streaming).
4. Tool Definition: Creating custom functions with specific decorators to define their name, description, and parameters.
5. Tool Integration: Registering tools with the SDK for autonomous agent use.
6. System Message Implementation: Utilizing a "system message" to define the agent’s personality and instructions.

Practical Applications Demonstrated

The demonstration showcased several applications:

• File Summarization: Loading and summarizing files using AI.
• Order Status Inquiry: An agent querying order status by ID, using a tool accessing a fictional database.
• Shipping Cost Calculation: A tool calculating shipping costs based on weight and region.
• Technical Support Assistant: Mentioned as a potential application interacting with GitHub and the file system.

Advanced Concepts: MCPs, Skills, and Model Selection

The discussion expanded on advanced concepts, including:

• MCPs (Model Connectors Protocol): Protocols for connecting agents to tools, facilitating interaction with external systems like PostgreSQL (via Neon DB) and file systems.
• Skills: “Skills” act as user manuals for tools (MCPs), reducing hallucinations and improving response accuracy.
• Model Flexibility: The SDK allows integration with various models and providers through BYOK. Sonnet 4.1 was highlighted as performing better than GPT-4.1 on certain instructions. Gemini’s large context window was also noted.

Streaming, Error Handling, and Context Management

The importance of streaming for real-time responses and error handling (subscribing/unsubscribing from sessions) was emphasized. Agents maintain conversational context, enabling complex, multi-turn interactions (e.g., tracking order status, asking about order components).

Cost Optimization & Agent Specialization

The SDK enables significant cost reduction compared to pre-built solutions by leveraging free tiers and usage-based pricing. Carlos stressed the importance of creating specialized agents focused on specific tasks, stating, “The more limited the people, the better they will behave. Don’t make a people that does everything.” He also cautioned, “Not all problems are solved with agents. In fact, if necessary, or if it’s not necessary, avoid agents.” Selecting the appropriate model for the task is also crucial, utilizing simpler, cheaper models when advanced capabilities aren’t required.

Example Implementations & Technical Details

Examples included an internal support agent answering questions about the GitHub API and Python automation, and an agent managing orders by querying fictional databases. The agent demonstrated writing to the file system, creating test.jam and robots.txt files. The example code was 227 lines, primarily tools and Python code, highlighting the simplicity of implementation.

Conclusion & Resources

The presentation concluded with a strong endorsement of the SDK’s flexibility, control, and cost-effectiveness. Resources were provided, including the GitHub repository containing the example code, Carlos’s YouTube channel for detailed tutorials, and information about Quicks’ corporate training services. The key takeaway is that the Copilot SDK empowers developers to build highly customized and efficient AI agents tailored to specific needs, offering a powerful alternative to generic, pre-packaged solutions.