Designing AI-Intensive Applications - swyx

Key Concepts

• AI Engineering evolution and standardization
• LLM OS (Large Language Model Operating System)
• LLM SDLC (Software Development Life Cycle)
• Building Effective Agents
• Human Input vs. AI Output ratio
• SPADE (Sync, Plan, Analyze, Deliver, Evaluate) model for AI-intensive applications

Conference Introduction and AI Engineering Evolution

The speaker opens the conference, addressing the audience and outlining the talk's purpose: to provide updates on AI engineering and the conference's structure. He humorously notes the last-minute registration rush, quantifying it with a "Genie coefficient for AIE organizer stress." He emphasizes the conference's role in tracking AI engineering's evolution, highlighting the doubling of tracks compared to the previous year. The conference aims to be more responsive than conferences like NeurIPS and more technical than conferences like TED, focusing on topics requested by attendees through surveys. The speaker encourages attendees to complete the survey to inform future conferences.

AI Engineering Innovation and Past Talks

The speaker highlights the conference's innovative spirit, mentioning being the first to have an MCP (Model Control Protocol) talk accepted by MCP. He acknowledges Sam Julian from Writer, Quinn and John from Daily, and Elizabeth Triken from Vappy for their contributions to the official chatbot and voice bot. He references his previous talks at the conference, noting the progression from defining the three types of AI engineers (2023) to the multidisciplinary nature of AI engineering (2024) and the focus on agent engineering (2025).

The "Emperor Has No Clothes" and the State of AI

The speaker notes that despite the advancements, AI engineering is still in its early stages, encouraging AI engineers to explore the "alpha to mine." He draws a parallel to the Solvay Conference of 1927, where foundational physics principles were established. The speaker poses the question: "What is the standard model in AI engineering?" He contrasts this with established standard models in other engineering fields like ETL, MVC, CRUD, and MapReduce. While acknowledging the existence of Retrieval-Augmented Generation (RAG), he questions its completeness as a standard model.

Candidate Standard Models in AI Engineering

The speaker presents several candidate standard models for AI engineering:

• LLM OS: An updated version of Karpavi's 2023 model, incorporating multimodality, standard tools, and MCP as the default protocol.
• LLM SDLC: Two versions of the LLM SDLC are mentioned, one with intersecting concerns of tooling. The key insight, derived from a conversation with Anker of BrainTrust, is that the early stages of the SDLC (LLM, monitoring, RAG) are becoming commoditized. The real value and hard engineering work lie in evaluations, security orchestration, and productionizing AI applications.
• Building Effective Agents: Referencing Barry's talk from the previous conference, the speaker acknowledges this as a widely accepted approach to building agents, although different definitions exist (e.g., OpenAI's). He mentions Dominic's improvements to the Agents SDK, building on OpenAI's swarm concept. The speaker's approach involves a descriptive, top-down model based on terms like intent, control flow, memory, planning, and tool use.

Human Input vs. AI Output and the SPADE Model

The speaker discusses the value delivered by AI, even in non-agentic systems, using AI News as an example. He asserts that the ratio of human input to valuable AI output is more relevant than arguing about the definition of agents. He presents a mental model tracking this ratio, ranging from copilot-style autocomplete to reasoning models and ambient agents with no human input.

He then describes the underlying process of AI News, which involves scraping, planning, summarizing, formatting, and evaluating. He generalizes this into the SPADE model for building AI-intensive applications:

1. Sync: Gather data from various sources.
2. Plan: Determine the processing steps.
3. Analyze: Parallel process and analyze the data.
4. Deliver: Present the results to the user.
5. Evaluate: Assess the quality of the output.

He notes that this model can incorporate knowledge graphs, structured outputs, and code generation (e.g., using Canvas or Cloud with artifacts).

Conclusion and Call to Action

The speaker concludes by emphasizing the importance of identifying new standard models for AI engineering that can improve applications and build valuable products. He encourages attendees to engage in discussions and explore potential standard models during the conference. He expresses his excitement for the conference and thanks the attendees.