AI Engineer World's Fair 2025 - Day 1 Keynotes & MCP track ft. Anthropic MCP team

AI Engineer World's Fair 2025 - Keynote Summaries

Key Concepts:

• AI Engineering Evolution
• Standard Models in AI Engineering (MOS, LM SDLC, Agent Building)
• Agentic Web
• Pair Programming vs. Peer Programming
• Software Factory vs. Agent Factory
• Signals Loop
• Graph RAG
• MCP (Model Context Protocol)
• Tool Calling
• Agent Economy
• Observability in AI Agents
• Strands Agents
• Vibe Coding
• AI Leapfrog Effect
• Execution as a Moat

Lorie Voss - Welcome and Introduction

• Main Topics and Key Points:
  • Welcoming attendees to the 2025 AI Engineer World's Fair.
  • Highlighting the impressive scale of the event and the AI revolution.
  • Introducing Llama Index as a framework for building AI applications.
  • Acknowledging the hype surrounding AI but emphasizing the real revolution happening.
  • Providing examples of real adoption: ChatGPT, GitHub Copilot, Azure AI.
  • Recognizing sponsors: Microsoft, AWS, Neo4j, Brain Trust, Graphite, Windinsurf, MongoDB, Daily, Augment Code, Work OS.
  • Celebrating the growth of the conference and the community.
• Important Examples, Case Studies, or Real-World Applications Discussed:
  • ChatGPT's rapid user adoption.
  • GitHub Copilot's subscriber base and integration with Microsoft 365.
  • Azure AI's $13 billion in annual revenue.
  • AWS's $87 billion investment in AI infrastructure.
• Key Arguments or Perspectives Presented, with Their Supporting Evidence:
  • AI is a real revolution, evidenced by widespread adoption and significant investments.
• Notable Quotes or Significant Statements with Proper Attribution:
  • Lorie Voss: "My name is Lori Voss i am VP of developer relations at Llama Index the best framework for building aic AI applications according to me."
  • Lorie Voss: "AGI will be achieved when models can say something actually funny."
• Logical Connections Between Different Sections and Ideas:
  • The introduction sets the stage for the conference by highlighting the importance of AI and the growth of the community.
  • The examples of real adoption provide evidence for the claim that AI is a real revolution.
  • The recognition of sponsors acknowledges the contributions of key players in the AI ecosystem.
• A Brief Synthesis/Conclusion of the Main Takeaways:
  • The AI revolution is real and is being driven by widespread adoption and significant investments.
  • The AI Engineer World's Fair is a testament to the growth of the AI community and the importance of building real-world applications.

Swix - State of AI Engineering

• Main Topics and Key Points:
  • Conference overview and agenda.
  • Evolution of AI engineering and its tracks.
  • Survey results on desired topics.
  • Innovation in conference experience (MCP, chatbots, voice bots).
  • Evolution of AI engineering focus (from types of engineers to multi-disciplinarity to agent engineering).
  • Emphasis on simplicity and avoiding over-complication.
  • Comparison of the current AI engineering moment to the Solvay Conference in physics.
  • Questioning the "standard model" in AI engineering.
  • Presenting candidate standard models: MOS, LM SDLC, Agent Building.
  • Highlighting the importance of human input vs. valuable AI output.
  • Introducing the SPA (Sync, Plan, Analyze) model for AI-intensive applications.
• Step-by-Step Processes, Methodologies, or Frameworks Explained:
  • MOS (Model-Orchestration-System): A standard model for AI application architecture.
  • LM SDLC (Software Development Life Cycle): A framework for developing AI applications, highlighting the increasing commoditization of early stages and the importance of evals and security.
  • Agent Building: Discussing different approaches to building effective agents, including the Enthropic model and a descriptive top-down model.
  • SPA (Sync, Plan, Analyze): A model for building AI-intensive applications involving scraping, planning, parallel processing, analysis, and delivery.
• Key Arguments or Perspectives Presented, with Their Supporting Evidence:
  • AI engineering is evolving and becoming more multi-disciplinary.
  • Simplicity and avoiding over-complication are key to success in AI engineering.
  • The focus should be on delivering value rather than arguing about terminology.
  • The standard model in AI engineering is still emerging.
• Notable Quotes or Significant Statements with Proper Attribution:
  • Swix: "We used to be low status people just deride GPT rappers and look at all the GPT rappers now all of you are rich."
  • Sumit Chintala (PyTorch lead): "AI news... it is not an agent."
• Technical Terms, Concepts, or Specialized Vocabulary with Brief Explanations:
  • Genie Coefficient: A measure of stress for the AI AIE organizer.
  • Graph RAG: Retrieval-Augmented Generation using graph databases.
  • MCP: Model Context Protocol, a standard for connecting AI models with external tools.
  • MOS: Model-Orchestration-System, a standard model for AI application architecture.
  • LM SDLC: Software Development Life Cycle for AI applications.
  • SPA: Sync, Plan, Analyze, a model for building AI-intensive applications.
• Logical Connections Between Different Sections and Ideas:
  • The talk begins with an overview of the conference and then transitions to a discussion of the evolution of AI engineering.
  • The presentation of candidate standard models builds on the earlier discussion of the need for a standard model in AI engineering.
  • The introduction of the SPA model provides a concrete example of a potential standard model.
• A Brief Synthesis/Conclusion of the Main Takeaways:
  • AI engineering is a rapidly evolving field with a growing need for standard models and best practices.
  • Simplicity and a focus on delivering value are key to success in AI engineering.
  • The standard model in AI engineering is still emerging, and there are several candidate models to consider.

Asha Sharma - The Open Agentic Web

• Main Topics and Key Points:
  • The evolution of AI models and the emergence of the agentic web.
  • The shift from pair programming to peer programming with AI.
  • The transition from a software factory to an agent factory.
  • The importance of a platform of AI-powered tools with trust and security.
  • The changing role of developers and the rise of AI-powered code maintenance.
  • The signals loop and the importance of fine-tuning models for personalization.
  • The changing infrastructure for building agentic applications.
  • The importance of open models and open protocols.
  • The need for agents to run locally and in the cloud.
• Important Examples, Case Studies, or Real-World Applications Discussed:
  • GitHub Copilot as a peer programmer.
  • FSY (Fuzzy Search Your Codebase) for AI-powered code maintenance.
  • Dragon, a healthcare co-pilot that automates scribing.
  • Gigapath, an open model for understanding pathology slides.
• Step-by-Step Processes, Methodologies, or Frameworks Explained:
  • Signals Loop: A continuous loop for improving AI models through fine-tuning and personalization.
• Key Arguments or Perspectives Presented, with Their Supporting Evidence:
  • The agentic web is creating new forces in AI engineering, including the shift to peer programming and the agent factory.
  • The signals loop is essential for improving the quality and accuracy of AI models.
  • Open models and open protocols are crucial for fostering innovation and collaboration in the AI ecosystem.
  • AI models need to run locally and in the cloud to meet the needs of diverse applications.
• Technical Terms, Concepts, or Specialized Vocabulary with Brief Explanations:
  • Agentic Web: A world in which agents interact with tools, models, and other agents.
  • Signals Loop: A continuous loop for improving AI models through fine-tuning and personalization.
  • Agentic RAG: Retrieval-Augmented Generation with iteration, evaluation, and planning.
  • A2A: Agent-to-Agent communication protocol.
  • MCP: Model Context Protocol, a standard for connecting AI models with external tools.
• Logical Connections Between Different Sections and Ideas:
  • The talk begins by outlining the evolution of AI models and then transitions to a discussion of the forces shaping AI engineering.
  • The presentation of the signals loop builds on the earlier discussion of the importance of fine-tuning models.
  • The discussion of open models and open protocols reinforces the theme of collaboration and innovation.
• A Brief Synthesis/Conclusion of the Main Takeaways:
  • The agentic web is transforming AI engineering and creating new opportunities for innovation.
  • The signals loop is essential for improving the quality and accuracy of AI models.
  • Open models and open protocols are crucial for fostering collaboration and innovation.
  • AI models need to run locally and in the cloud to meet the needs of diverse applications.

Sarah Guo - Cursor for X

• Main Topics and Key Points:
  • The capabilities of AI, including reasoning, agents, and multimodality.
  • The increasing competitiveness of the model market.
  • The success of Cursor and the opportunity to build "Cursor for X" in other industries.
  • The importance of domain knowledge and workflow knowledge in building AI applications.
  • The AI leapfrog effect in conservative industries.
  • The continued relevance of co-pilots and augmentation.
  • The importance of execution as a moat in AI.
• Important Examples, Case Studies, or Real-World Applications Discussed:
  • Cursor as a successful AI-powered coding tool.
  • Cognition as a top committer in many companies.
  • Windsurf's acquisition by OpenAI.
  • Sierra, Harvey, and Open Evidence as examples of AI adoption in conservative industries.
• Key Arguments or Perspectives Presented, with Their Supporting Evidence:
  • The model market is becoming more competitive, with new players emerging and prices falling.
  • The opportunity to build value around LLMs exists in every vertical and profession.
  • The most conservative industries are adopting AI fastest.
  • Co-pilots are still really underrated and offer a path of least frustration.
  • Execution is the moat in AI, not just the model.
• Logical Connections Between Different Sections and Ideas:
  • The talk begins by outlining the capabilities of AI and then transitions to a discussion of the application layer.
  • The presentation of Cursor as a successful AI-powered coding tool leads to the idea of building "Cursor for X" in other industries.
  • The discussion of the AI leapfrog effect provides evidence for the claim that the opportunity to build value around LLMs exists in every vertical and profession.
• A Brief Synthesis/Conclusion of the Main Takeaways:
  • The opportunity to build value around LLMs exists in every vertical and profession.
  • Domain knowledge and workflow knowledge are crucial for building successful AI applications.
  • The most conservative industries are adopting AI fastest.
  • Co-pilots are still really underrated and offer a path of least frustration.
  • Execution is the moat in AI, not just the model.

Simon Willison - State of LLMs

• Main Topics and Key Points:
  • Review of significant model releases in the past six months.
  • Challenges in evaluating model quality and the limitations of benchmarks.
  • Introduction of a personal benchmark: generating an SVG of a pelican riding a bicycle.
  • Discussion of various models: AWS Nova, Llama 3, Deepseek, Mistral, Claude 3, GPT-4.5, 01 Pro, Gemini 2.5 Pro, GPT40.
  • Analysis of bugs in LLMs and their implications.
  • Emphasis on the importance of tools and reasoning.
  • Discussion of the lethal trifecta: private data, malicious instructions, and exfiltration.
• Important Examples, Case Studies, or Real-World Applications Discussed:
  • The use of a personal benchmark to evaluate model quality.
  • The bugs in ChatGPT and Claude 4 and their implications for AI safety.
• Key Arguments or Perspectives Presented, with Their Supporting Evidence:
  • Local models are now good and worth paying attention to.
  • The prices of good models have crashed significantly.
  • Tools and reasoning are essential for building powerful AI systems.
  • The lethal trifecta poses a significant risk to AI security.
• Technical Terms, Concepts, or Specialized Vocabulary with Brief Explanations:
  • Pelican on a Bicycle: A personal benchmark for evaluating model quality.
  • Lethal Trifecta: The combination of private data, malicious instructions, and exfiltration that poses a significant risk to AI security.
• Logical Connections Between Different Sections and Ideas:
  • The talk begins with a review of recent model releases and then transitions to a discussion of the challenges in evaluating model quality.
  • The presentation of the personal benchmark provides a concrete example of how to evaluate model quality.
  • The discussion of bugs in LLMs leads to a discussion of AI safety and the lethal trifecta.
• A Brief Synthesis/Conclusion of the Main Takeaways:
  • The field of LLMs is rapidly evolving, with new models being released at a rapid pace.
  • Evaluating model quality is challenging, and personal benchmarks can be helpful.
  • Tools and reasoning are essential for building powerful AI systems.
  • AI safety is a growing concern, and the lethal trifecta poses a significant risk.

Steven Chin and Andreas Kleger - Agentic Graph RAG

• Main Topics and Key Points:
  • Graph RAG and its ability to solve sci-fi memes.
  • Demo of end-to-end importing data and building a knowledge graph.
  • Announcement of the new Neo4j startup program.
• Important Examples, Case Studies, or Real-World Applications Discussed:
  • The use of graph RAG to connect the Terminator and Blade Runner through the theme of technoir.
• A Brief Synthesis/Conclusion of the Main Takeaways:
  • Graph RAG is a powerful tool for connecting unstructured data and enabling intelligent queries.
  • The Neo4j startup program provides a free way for startups to build AI applications with graph technology.

Antia Bar - Building Agents at Cloud Scale

• Main Topics and Key Points:
  • The transformation of customer experiences with AI agents.
  • Examples of AI agents at Amazon: Alexa and Amazon Q developer.
  • The model-driven approach to building agents.
  • Introduction of Strands Agents, an open-source Python SDK.
  • Integration of MCP with Strands Agents.
  • Deployment of MCP servers as Lambda functions.
  • The future of agentic interactions and the agent economy.
• Important Examples, Case Studies, or Real-World Applications Discussed:
  • Alexa's reimagining with AI agents.
  • Amazon Q developer agent for the CLI.
• Step-by-Step Processes, Methodologies, or Frameworks Explained:
  • Model-Driven Approach: A method for building agents that relies on LLMs to reason, plan, and take action.
• Key Arguments or Perspectives Presented, with Their Supporting Evidence:
  • AI agents will transform customer experiences and create new opportunities for innovation.
  • The model-driven approach simplifies the development of AI agents.
  • Open protocols like MCP are essential for enabling agentic interactions.
• Technical Terms, Concepts, or Specialized Vocabulary with Brief Explanations:
  • Strands Agents: An open-source Python SDK for building AI agents.
• Logical Connections Between Different Sections and Ideas:
  • The talk begins by outlining the transformation of customer experiences with AI agents and then transitions to a discussion of how to build these agents.
  • The presentation of Strands Agents provides a concrete example of a tool for building AI agents.
  • The discussion of MCP and open protocols reinforces the theme of collaboration and innovation.
• A Brief Synthesis/Conclusion of the Main Takeaways:
  • AI agents will transform customer experiences and create new opportunities for innovation.
  • The model-driven approach simplifies the development of AI agents.
  • Open protocols like MCP are essential for enabling agentic interactions.

Kevin Ho - What's Next for Agentic IDEs

• Main Topics and Key Points:
  • The shared timeline between human and AI as the secret sauce of Windsurf.
  • The need for Windsurf to be everywhere and ingest context from every source.
  • The ability for Windsurf to do and write everything, not just code.
  • The development of a new software engineering model called SU1.
  • The importance of a data flywheel for improving AI products.
• Key Arguments or Perspectives Presented, with Their Supporting Evidence:
  • The shared timeline between human and AI is essential for creating a seamless and intuitive user experience.
  • Windsurf needs to be everywhere and ingest context from every source to be truly effective.
  • The future of software engineering is not just about code generation but about a broader set of tasks.
  • A data flywheel is essential for continuously improving AI products.
• Logical Connections Between Different Sections and Ideas:
  • The talk begins by outlining the shared timeline between human and AI and then transitions to a discussion of the need for Windsurf to be everywhere and do everything.
  • The presentation of SU1 provides a concrete example of a new software engineering model.
  • The discussion of the data flywheel reinforces the theme of continuous improvement.
• A Brief Synthesis/Conclusion of the Main Takeaways:
  • The shared timeline between human and AI is essential for creating a seamless and intuitive user experience.
  • Windsurf needs to be everywhere and ingest context from every source to be truly effective.
  • The future of software engineering is not just about code generation but about a broader set of tasks.
  • A data flywheel is essential for continuously improving AI products.

Greg Brockman - Fireside Chat

• Main Topics and Key Points:
  • Greg's early experiences with coding and his decision to drop out of MIT to join Stripe.
  • The challenges of building Stripe in its early days.
  • The importance of independent study and self-directed learning.
  • The evolution of AI and the factors that convinced Greg that AGI was possible.
  • The relationship between engineering and research at OpenAI.
  • The challenges of scaling AI systems and the importance of reliability.
  • The future of AI infrastructure and the need for both long compute and real-time capabilities.
  • The importance of algorithms in the future of AI.
  • The potential for AI to transform the economy and create new opportunities for innovation.
• Key Arguments or Perspectives Presented, with Their Supporting Evidence:
  • The best way to learn is to just build things and experience things out in the world.
  • The relationship between engineering and research is essential for success in AI.
  • Algorithms are becoming increasingly important in the future of AI.
  • AI has the potential to transform the economy and create new opportunities for innovation.
• Logical Connections Between Different Sections and Ideas:
  • The conversation begins with Greg's early experiences and then transitions to a discussion of his work at Stripe and OpenAI.
  • The discussion of the relationship between engineering and research builds on the earlier discussion of the importance of independent study and self-directed learning.
  • The discussion of the future of AI infrastructure and the importance of algorithms leads to a discussion of the potential for AI to transform the economy.
• A Brief Synthesis/Conclusion of the Main Takeaways:
  • The best way to learn is to just build things and experience things out in the world.
  • The relationship between engineering and research is essential for success in AI.
  • Algorithms are becoming increasingly important in the future of AI.
  • AI has the potential to transform the economy and create new opportunities for innovation.