Building Applications with AI Agents — Michael Albada, Microsoft

Summary

In this insightful presentation, Michael Alba, a principal applied scientist at Microsoft, explores the exciting yet challenging landscape of building applications with AI agents, particularly in cybersecurity. Drawing from his experience with Security Copilot and machine learning at Uber, he discusses the evolution of agentic systems, their potential, and common obstacles. Key topics include balancing agency and effectiveness, optimal tool use, and AI orchestration patterns. He emphasizes the importance of rigorous evaluation, the need for security in AI systems, and the future of multi-agent protocols. His talk highlights the potential of agentic AI to revolutionize productivity and problem-solving across industries.

Takeaways

• 😀 The rise of agentic systems is creating a significant shift in the field of AI, with a 254% increase in agent-focused companies at Y Combinator over the past three years.
• 😀 Building agentic systems involves more than just improving accuracy; it's about balancing agency with effectiveness to avoid compromising performance.
• 😀 Robotic process automation (RPA) serves as an example of low agency but high efficacy, showing that effective automation doesn't always require high levels of flexibility.
• 😀 Tool use in agentic systems is powerful, allowing foundation models to call external functions via APIs, but it requires careful management to avoid confusion and reduce semantic overlap.
• 😀 Keeping the orchestration of agent tasks simple is crucial for maintaining reliability, reducing cost, and ensuring ease of maintenance.
• 😀 Multi-agent systems are an effective way to scale, especially as the number of tools grows. This approach involves splitting tasks into semantically similar groups to reduce overload on a single agent.
• 😀 Evaluation is key: a rigorous, test-driven approach is essential for improving agent performance and making decisions on the best configuration of agents, tools, and models.
• 😀 Synthetic data generation tools like Intel Agent can help augment evaluation sets when raw user data isn't available, assisting in the testing of agent systems before their deployment.
• 😀 Observability is a significant challenge, as AI agents can generate diverse outcomes. Using tools like OpenTelemetry is vital to track logs and trace failures to improve system reliability.
• 😀 Common pitfalls in agentic system development include insufficient evaluation, overly complex tools, excessive complexity, and the lack of a tight learning loop.
• 😀 Security concerns are growing with agentic systems, especially in the cyber security domain. It's crucial to design for safety at every layer and implement trip wires and detectors to allow for human oversight.

Q & A

What is the main focus of Michael Alba's presentation?

Michael Alba's presentation focuses on building applications with AI agents, specifically discussing the challenges, components, and design principles involved in developing agentic systems in AI.

What experience does Michael Alba have in the field of AI and security?

Michael Alba has been a Principal Applied Scientist at Microsoft for two years, contributing to the development of Security Copilot and AI agents in the cybersecurity division. Before that, he worked at Uber on machine learning for four years and in startups.

What is the significance of AI agents in modern technology, according to the speaker?

AI agents are significant because they can reason, act, communicate, and adapt to solve complex tasks. These agents represent a shift toward more flexible, adaptable systems that can respond to dynamic inputs, unlike traditional fixed automations like robotic process automation.

What does the term 'agentic' refer to in the context of AI systems?

In the context of AI systems, 'agentic' refers to systems that exhibit behaviors such as reasoning, acting, communicating, and adapting. These systems are not merely task automations but are capable of complex decision-making and adapting to changing scenarios.

Why is it challenging to move from prototype AI systems to complex, real-world applications?

Moving from prototype AI systems to real-world applications is challenging because initial prototypes may achieve 70% accuracy, but scaling to handle more complex, dynamic tasks often requires overcoming significant hurdles, such as maintaining high performance across diverse use cases and inputs.

How does tool use enhance the functionality of AI agents?

Tool use enhances AI agents by allowing them to invoke functions and access external APIs, thereby expanding their ability to perform a wide range of tasks. This makes the agent more capable but also introduces risks that need careful consideration to ensure functionality is exposed responsibly.

What is the recommendation regarding exposing tools to AI agents?

The recommendation is to avoid exposing too many tools to a single AI agent at once. Too many tools can confuse the agent and decrease accuracy. It's better to group related tools logically and expose them only when necessary to maintain the agent's effectiveness.

What are the advantages of using simple workflow patterns in AI agent development?

Simple workflow patterns, such as single chains or basic branching logic, help maintain clarity, ease of measurement, and reduce complexity in AI systems. These patterns also improve reliability and minimize costs, making it easier to scale and deliver consistent value to customers.

What is the role of evaluation in AI agent development?

Evaluation is critical in AI agent development to assess the effectiveness of models and workflows. By rigorously evaluating performance, teams can identify issues, adjust hyperparameters, and improve the system iteratively, ultimately enhancing the agent’s effectiveness and reliability.

What are some of the common pitfalls in building AI agents, as mentioned by Michael Alba?

Some common pitfalls include insufficient evaluation, exposing too many tools to the agent, excessive complexity in agentic systems, and failure to maintain a tight learning loop. Additionally, it's crucial to design systems for safety and ensure agents are resilient to potential vulnerabilities.