From: aidotengineer
The AI Engineer Summit has pivoted to focus on agent engineering, indicating a significant shift in the industry’s direction [00:02:44]. This strategic decision involves saying “no” to other areas like Retrieval Augmented Generation (RAG), open models, and GPUs, to concentrate solely on the domain of agents [00:02:53]. This focus is partly driven by the high engagement with “agentic” talks on YouTube from previous years [00:03:17].
Defining an Agent
For any conference on agents, defining the term “agent” is a necessary public service [00:05:25]. From various perspectives:
- Machine Learning (ML): Agents are viewed in the context of reinforcement learning environments, focusing on actions to achieve goals [00:05:41].
- Software Engineers: May reductively define an agent as “a for loop” [00:05:51].
Simon Willison, considered a “patron saint” in the AI engineering community, has crowdsourced over 300 definitions of what an agent is [00:06:06]. Common themes among these definitions include:
- Goal orientation [00:06:17]
- Tool use [00:06:19]
- Control flow [00:06:20]
- Long-running processes [00:06:21]
- Delegated authority [00:06:22]
- Small, multi-step task completion [00:06:24]
OpenAI also introduced a new definition for agents, which is an important development to monitor [00:06:52].
Why Agents are Gaining Traction Now
The question arises: why are agents effective now, when they weren’t a year or two ago [00:07:12]? Several factors contribute to their current viability:
- Improved Capabilities: Models demonstrate better reasoning and tool use [00:07:39]. Capabilities have been steadily growing and are hitting human baselines around the present time [00:07:27].
- Model Diversity: OpenAI’s market share has decreased from approximately 95% two years ago to 50%, leading to a more diverse landscape of models [00:07:51]. The emergence of new Frontier Model Labs challenges OpenAI’s dominance [00:08:02].
- Lower Cost of Intelligence: The cost of GPT-4 level intelligence has decreased by 1,000 times in the last 18 months [00:08:16]. Similar cost reductions are beginning for O1-level intelligence [00:08:23].
- RL Fine-tuning Options: The availability of reinforcement learning (RL) fine-tuning options is also a contributing factor [00:08:28].
- Multi-agents and Faster Inference: Advancements in multi-agent systems and faster inference capabilities, supported by improved hardware, are also playing a role [00:08:49].
Impact and Future Potential
The saying “2025 is the year of Agents” is a common prediction by industry leaders like Satya Nadella, Roman, Greg Brockman, and Sam Altman [00:04:20]. While there has been skepticism and fatigue around the term “agents” [00:04:48], OpenAI’s shifting stance—from advising against “agents” in branding to now recommending it—signals a growing belief in their importance [00:05:08].
Agent Use Cases and Anti-Use Cases
Agent technology shows promise in specific areas:
- Product-Market Fit (PMF): Coding agents and support agents currently have PMF [00:09:12]. Deep research agents are also gaining PMF [00:09:15].
- Emerging Use Cases: “Everything plus agent works,” such as agent + RAG, agent + sentiment analysis, and agent + search, are seen as simple formulas for financial success in 2025 [00:04:00].
However, there are also “anti-use cases” for agents:
- Booking flights [00:09:25]
- Booking Instacart orders [00:09:34]
Growth Trajectory and Future of AI Engineering
OpenAI reported 400 million users, a 33% growth in three months [00:09:47]. The growth of ChatGPT, particularly after a period of stagnation, is tightly linked to the shipment of agentic models [00:10:09]. Agentic models have doubled ChatGPT’s usage [00:10:24]. Projecting this trend, ChatGPT is expected to reach one billion users by the end of the current year, quintupling its user base from September of the previous year [00:10:28].
The growth of any AI product is expected to be closely tied to its reasoning capabilities and the number of agents it can provide to users [00:10:41]. This massive scale suggests that by year-end, one-eighth of the world’s population will be using ChatGPT [00:10:49].
The job of an AI engineer is evolving towards building agents, similar to how Machine Learning Engineers build models and Software Engineers build software [00:11:00]. This highlights the transformative potential and impact of agents in the AI landscape.