Why Agent Engineering
swyx landmark keynote on why 2025 is the year of agent engineering. Learn the 6 enabling factors, agent definitions, PMF use cases, ChatGPT growth analysis, and why agents matter now when they didn't work a year ago.
"I think that the job of AI engineers is now evolving towards building agents in the same way that ML engineers build models and software engineers build software."
— swyx, AI Engineer Summit Lead (18:20)
Watch (18:20)Enabling Factors
Agent Definitions
PMF Use Cases
Growth Stalled → Doubled
Executive Summary
swyx delivers a landmark keynote defining agent engineering as a distinct discipline for 2025. He explains why agents are working now when they failed a year ago, documenting six converging factors: better reasoning models, improved tool use, model diversity, dramatically lower costs (1000x cheaper for GPT-4-level intelligence), RL fine-tuning options, and multi-agent systems.
The talk presents six definitions of agents from different perspectives (goals, tools, control flow, long-running processes, delegated authority, multi-step task completion), emphasizing that the field is still forming its identity. swyx reveals that AI Engineer Summit deliberately pivoted to focus exclusively on agent engineering this year, saying no to RAG, open models, and GPUs to go deep on one theme.
The most compelling insight comes from ChatGPT growth analysis. ChatGPT spent a full year with zero growth because they didn't ship any agentic models. When o1 models launched, usage doubled overnight. swyx projects ChatGPT will hit 1 billion users by end of 2025 (1/8 of world population), proving that growth is tightly coupled to reasoning capabilities and agentic features.
On product-market fit, swyx identifies three clear winners: coding agents, support agents, and deep research agents. He calls out anti-patterns that should be retired: flight booking agents and Instacart ordering agents. The data shows that AI engineering is emerging as its own discipline, distinct from both ML engineering (which sees it as "MLE plus prompts") and software engineering (which sees it as "just calling LLM APIs").
The State of AI Engineering
swyx has been documenting the evolution of AI engineering since the first AI Engineer Summit. The discipline is maturing and spreading, but faces identity resistance from both ML engineers and software engineers who underestimate its distinct nature.
ML Engineers' View
"AI engineer is mostly an ML engineer plus a few prompts."
Language clue: "Test time compute" - the only reason to run inference is to test it.
Software Engineers' View
"AI engineer is mostly software engineering and calling a few LLM APIs."
Language clue: "Reasoning" - software engineers talk about reasoning systems.
AI Engineering is Its Own Discipline
swyx: "I still say things like 'AI engineering is 90% software engineering, 10% AI.' I think that will grow over time." The discipline is emerging as distinct from both traditional ML engineering and software engineering, with its own patterns, language, and practices. This is the year when AI engineering starts to spread out and establish its identity.
Six Definitions of Agents
Before any agent engineering conference, swyx argues we must define what we mean by "agent." He presents six definitions from Simon Willison's crowdsourced 300+ definitions, plus OpenAI's new official definition.
"I have one slide to define agents. I could do it in one slide."
— swyx - On the challenge of defining agents
11:40Goals
Agents achieve specific objectives through autonomous decision-making.
Tools
Agents use external tools and APIs to interact with the world.
Control Flow
Agents manage their own execution flow with loops and conditionals.
Long-Running Processes
Agents operate over extended time periods, not single-turn interactions.
Delegated Authority
Agents make decisions on behalf of users with varying autonomy levels.
Multi-Step Task Completion
Agents break down complex tasks into sequential steps.
OpenAI's New Agent Definition
During the summit, OpenAI dropped a new agent definition on the live stream. swyx emphasizes paying attention to this because OpenAI is building products on top of this definition. The definition represents their official stance on what constitutes an agent.
Watch OpenAI's Definition (14:10)Key Takeaways for AI Engineers
Six Factors Enable Agents in 2025
- •Better reasoning models hitting human baselines
- •Improved tool use and structured outputs
- •Better tools and infrastructure (MCP, frameworks)
- •Model diversity - OpenAI down to 50% market share
- •Cost of intelligence down 1000x in 18 months
- •RL fine-tuning and multi-agent systems
Agents Drive User Growth
- •ChatGPT stagnated for 1 year without agentic models
- •o1 models doubled ChatGPT usage overnight
- •1B users projected by end of 2025 (1/8 world population)
- •Growth tightly coupled to reasoning and agentic features
- •Massive money left on the table for agentic products
Clear PMF in 3 Categories
- •Coding agents - strong product-market fit
- •Support agents - validated in production
- •Deep research agents - OpenAI's 400M users prove it
- •Anti-patterns: flight booking, Instacart, astroturfing
AI Engineering is Its Own Discipline
- •Distinct from ML engineering (not just 'MLE + prompts')
- •Distinct from software engineering (not just 'calling LLM APIs')
- •Will grow from 90% SE / 10% AI to more balanced split
- •Job of AI engineers: build agents like MLEs build models
Everything + Agent = Money in 2025
- •Agent + RAG works
- •Agent + search works
- •Agent + [any domain] works
- •Simple formula for AI product success
Six Agent Definitions Exist
- •Goals - autonomous objective achievement
- •Tools - using external APIs and functions
- •Control flow - managing own execution
- •Long-running - extended time operations
- •Delegated authority - making decisions for users
- •Multi-step task completion - complex task decomposition
Source Video
Why Agent Engineering
swyx (Shawn Wang) • AI Engineer Summit Lead
Research Note: This comprehensive analysis is based on swyx's landmark keynote at AI Engineer Summit 2025. All quotes are timestamped and link to exact moments in the video for validation. The analysis covers the six enabling factors for agents, six definitions of agents, PMF use cases (coding, support, deep research), ChatGPT growth correlation with agentic models, and the evolution of AI engineering as a distinct discipline.
Key Concepts: Agent engineering, swyx, AI Engineer Summit, agents 2025, agent PMF, ChatGPT growth, o1 models, model diversity, cost of intelligence, RL fine-tuning, multi-agent systems, agent definitions, AI engineering discipline, agent use cases, product-market fit