AI engineering is becoming its own discipline as the work moves beyond software engineering with a few LLM calls or machine learning with a few prompts.
2
Agents are more practical now because models have better reasoning and tool use, model choices have multiplied, intelligence costs have fallen, and inference has become faster.
3
The strongest agent use cases include coding, support, and deep research, while flight booking and grocery ordering demos do not show useful product demand.
Summary
Shawn Wang explains why AI Engineer Summit has shifted its focus toward agent engineering in 2025. He argues that AI engineering is still early and sits between machine learning and software engineering, but it is developing its own language, practices, and identity. Wang defines agents through ideas such as goals, tools, control flow, long-running processes, delegated authority, and multi-step tasks. He then gives several reasons agents are becoming practical: better reasoning, better tool use, more model choice, sharply lower intelligence costs, reinforcement learning fine-tuning, multi-agent work, and faster inference. He connects agent adoption to product growth, pointing to coding agents, support agents, and deep research as use cases with product-market fit. He is also direct about weak demos, especially agents that book flights or place grocery orders. His central claim is that AI engineers are increasingly building agents in the same way machine learning engineers build models and software engineers build software.
AI engineering is developing into its own discipline
Wang describes resistance from both sides of the field. Machine learning engineers may see an AI engineer as mostly an MLE with a few prompts, while software engineers may see the role as software engineering with a few LLM API calls. He expects AI engineering to become its own discipline, although it is still very early. He previously described the work as "90% software engineering, 10% AI," but says the AI portion will grow. He also treats the field as an anthropology of how people describe themselves, form groups, and build identities. Different communities use different language, such as "inference time compute" for AI engineers and "reasoning" for software engineers.
Agent engineering narrows the conference focus while opening other combinations
Wang says the summit's move toward agent engineering is not a rejection of retrieval-augmented generation, open models, or GPUs. Closing some doors made room for other topics. He found that "everything plus agent" works as a broad formula, including agents combined with RAG, computer vision, and search. The summit also changed its speaker policy after viewers asked who was putting agents into production. Instead of more vendor pitches from people who build agent frameworks, the organizers asked practitioners to share production experience. Wang says this made curation harder because those speakers had little direct incentive to appear on stage and disclose their work.
Wang says every agent conference needs to define "agent," then presents a collection of definitions associated with Simon Willison's crowdsourced work. The definitions focus on goals, tools, control flow, long-running processes, delegated authority, and small multi-step task completion. He contrasts this with machine learning discussions about reinforcement-learning environments and actions that achieve goals. Software engineers may reduce an agent to a for loop. Wang does not insist on one universal definition. Instead, he suggests checking whether a proposed definition passes several properties that people commonly associate with agents. He also points attendees to OpenAI's new agent definition, which had been released during the summit.
Agents are becoming practical because several conditions changed at once
Wang asks why agents are working now when they did not work a year or two earlier. Better reasoning, better tool use, and better tools are part of the answer. He also points to model diversity, with OpenAI's market share falling from roughly 95 percent two years earlier to about 50 percent. The cost of GPT-4-level intelligence, in his estimate, fell by 1,000 times over the previous 18 months. Reinforcement-learning fine-tuning options, outcome-based pricing, multi-agent research, faster inference, and improved hardware add to the change. His explanation is cumulative. Agents became more viable because capability, choice, price, and speed all moved in their favor.
Coding, support, and deep research have stronger demand than flight booking
Wang groups coding agents and support agents with talks about product-market fit, and says deep research can now be added to that group. He also names use cases that should stop being used as agent showcases. He asks people to stop demoing flight-booking agents because he wants to book his own flights. He makes the same point about Instacart orders. These examples are a test of whether an agent solves a problem users actually want to delegate. A technically impressive workflow is not enough if the user prefers to keep control of the task.
Agent capabilities are tied to the growth of AI products
Wang cites OpenAI's report of 400 million users, describing it as 33 percent growth in three months. He connects ChatGPT's earlier period of flat growth with the lack of agentic model releases during that time. In his reading of the usage chart, the o1 models doubled ChatGPT usage. He predicts that ChatGPT could reach one billion users by the end of 2025, while acknowledging that this is a projection from the chart. His broader point is that growth in AI products will be closely tied to improvements in reasoning and to the number of agents that products can ship for users.
The AI engineer's work is moving toward building agents
Near the end, Wang states that the job of the AI engineer is evolving toward building agents. He compares this with machine learning engineers building models and software engineers building software. The talk is an introduction to the summit's sessions rather than a detailed implementation guide. Its practical message is about where to spend attention: understand the emerging agent vocabulary, examine why the current generation works better, and focus on applications with real user demand. Wang closes by welcoming the audience to the event and introducing the speakers who will discuss those applications in more detail.
"The job of AI is now evolving towards building agents in the same way that MLEs build models, software engineers build software."11:00
Who should watch
You are deciding whether your team should describe its work as AI engineering, machine learning, software engineering, or agent engineering.
You want a compact explanation of why agents became more practical in 2025, including changes in model cost, capability, choice, and speed.
You are evaluating agent product ideas and need a blunt filter for separating useful delegation from demos such as flight booking.
