AI Engineer Summit 2025

Why Agent Engineering

swyx (Shawn Wang)11:45

Loading slides…

Read the talk

Why Agent Engineering

As models improve and inference gets cheaper, AI engineering is moving toward a more specific deliverable: agents that do useful work for users.

From a talk by swyx (Shawn Wang)

What does an AI engineer build?

How is AI engineering doing? There is now an O’Reilly book: Chip Huyen’s AI Engineering. There is also a hype-cycle chart placing the field near the peak—a setup for swyx’s joke that it can only go downhill from here. Recognition and inflated expectations have arrived together.

Slide showing Chip Huyen holding the AI Engineering book beside a post with a hype-cycle chart marking AI Engineering near the peak.Open full source frame
How is AI Engineering doing? A book launch alongside a hype-cycle chart.

The earlier task was to establish that the discipline existed. The Rise of the AI Engineer named the emerging role; the first AI Engineer Summit explored three types of AI engineer; the following World’s Fair examined its maturation across disciplines. Those frameworks helped people organize teams, but a label that has gained acceptance still needs a distinct account of the work.

Two neighboring professions offer competing descriptions. From machine learning engineering, AI engineering can look like machine learning plus a few prompts. From software engineering, it can look like software plus a few LLM API calls. Neither description leaves much room for an independent discipline. Swyx’s own current heuristic is 90% software engineering and 10% AI, but he expects the AI portion to grow as the work develops its own practices.

Professional language already reveals different priorities. A machine learning engineer might say test-time compute, while an AI engineer says inference-time compute because inference is the product’s everyday work. A software engineer might simply call it reasoning. This is an anthropological view of engineering: people describe their work, form groups around shared concerns, and eventually establish industries. Agent engineering is the next proposed focus for that process.

0:340:48
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:34 · section reference included

From frameworks to production

Making the Summit an agent engineering conference required excluding RAG, open models, and GPUs as standalone topics. That was a choice about the program’s scope, not a rejection of those technologies. Narrowing the subject opened space for more specific kinds of agent work, and the previous year’s strongest-performing YouTube talks suggested that the audience wanted that focus.

Audience interest alone did not solve the selection problem. The available speakers were largely people who built agent frameworks for a living, while the unanswered question was who actually used agents in production. The new rule was no vendor pitches. Recruiting practitioners made curation harder: people running systems have less incentive to appear onstage than people selling the tools used to build them. Their experience was precisely what the conference needed.

Agents can extend existing application categories. Agent plus RAG, agent plus code generation, and agent plus search were swyx’s compact formula for commercial opportunity in 2025. The accompanying slide expands the combinations to vision, voice, wearables, finance, and money. An agent focus therefore need not isolate the field from the rest of AI engineering; it changes what builders try to accomplish with those components.

Slide titled “have you tried rubbing an agent on it?” lists eight agent combinations and examples including Agentic RAG, Deep Research, Computer Use Agent, Bee, and Solana/Stripe.Open full source frame
Agents combined with RAG, code generation, search, vision, voice, wearables, finance, and money.
2:483:03
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

2:48 · section reference included

Reconsidering the year of agents

Calling 2025 the year of agents does not make it so. Predictions often confuse what their authors want with what will happen. Swyx names Satya Nadella, Roman, Greg Brockman, and Sam Altman as advocates of that expectation, while acknowledging that he and his Latent.Space co-host Alessio had been skeptical themselves.

The audience had its own fatigue: in the previous day’s Family Feud exercise, agents ranked second among buzzwords people were tired of hearing. In a March 2024 conversation with David Luan, swyx and Alessio had advised people to remove agents from their branding. Now swyx was advising them to put it back. The reversal sets up two questions: what counts as an agent, and what has changed enough to make one useful?

4:184:31
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:18 · section reference included

Several definitions, no settled boundary

The definition depends partly on where the engineer starts. Machine learning brings reinforcement learning environments, actions, and goals. Software engineering offers the deliberately reductive joke: put it in a for loop. AI engineering has not yet settled on its own preferred boundary. These perspectives emphasize different parts of a system, even when they describe the same application.

Simon Willison’s crowdsourcing gathered hundreds of responses about what an agent is. Swyx organizes the recurring definitions around six properties:

EmphasisDefining concern
GoalsAutonomous pursuit of an objective
ToolsAn LLM using tools
Control flowAn LLM directing what happens next
Long-running processesWork that persists over time
Delegated authorityPermission to act on someone’s behalf
Multistep completionCarrying a task through several steps

These are different criteria, not a single checklist that every definition requires.

The full slide also includes metaphors, reductive descriptions, and specific examples. That last approach offers a practical test: collect systems people broadly agree are agents, then check whether a proposed definition includes them. It grounds a terminology debate in recognizable cases.

Three-column slide beneath a Simon Willison post lists autonomous goal-oriented systems, tool-using LLMs, LLM-driven control flow, long-running processes, delegated authority, multi-step task completion, metaphors, reductiveness, and specific examples.Open full source frame
Nine ways of describing agents, from goal-oriented systems to specific examples.

Definitions also matter because they shape platforms. OpenAI had introduced another agent definition in the previous day’s presentation, and swyx urges attention to it because the company will build around that understanding. The significance here is the connection between a definition and the systems it encourages, rather than a final resolution of the terminology.

5:205:29
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:20 · section reference included

Why agents may work now

Why should agents work now when they struggled a year or two earlier? The first answer is capability. Swyx presents trajectories from 2023 to 2025 approaching human baselines. Without named tasks and evaluation conditions, that is a directional argument about improving models, not a claim that agents have reached human performance across arbitrary work.

He separates three familiar contributors:

  • Better reasoning: stronger model capability for working through a task.
  • Better tool use: improved ability to use external capabilities.
  • Better tools: a stronger ecosystem for connecting models to useful functions and information, including the Model Context Protocol, or MCP.

The distinction matters: a model’s ability to use a tool and the availability of a useful tool are separate parts of the engineering problem.

Model diversity adds another source of opportunity. Swyx estimates that OpenAI’s share fell from roughly 95% two years earlier to 50%, without specifying the market denominator. He also points to two possible frontier-lab challengers emerging in the preceding week. The argument is that builders face a more varied competitive landscape; which providers will prevail remains uncertain.

7:087:18
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:08 · section reference included

Cheaper intelligence, more operating options

Capability is only part of viability; the cost of using it matters too. Swyx calls the decline in intelligence costs “super Moore’s law.” He claims that the cost of GPT-4-level intelligence fell 1,000-fold over 18 months, without specifying the benchmark or pricing basis. He sees a similar curve beginning for o1-level intelligence. The proposed enabling change is that an application can afford intelligence that would previously have been too expensive.

Reinforcement learning fine-tuning introduces another option for builders. Swyx explicitly says he has no personal experience with it and points to Will’s later presentation for a practitioner’s account. It belongs in the list of emerging capabilities, rather than as a technique demonstrated in this keynote.

The remaining changes concern both delivery and economics. A conversation with Brett Taylor raised charging for outcomes instead of costs, shifting the commercial unit toward the result a customer wants. Work on multiple agents expands how systems can organize tasks, while better hardware enables faster inference. Swyx points to Latent.Space conversations for the supporting discussions; here, these are additional contributors to agent viability, not detailed implementation recipes.

8:098:17
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:09 · section reference included

Useful delegation and unwanted delegation

Where does that combination already create value? Referring to Barry Zhang’s Building effective agents, swyx identifies coding and support agents as having product-market fit. He adds deep research to that category and anticipates further use cases in the conference program. These are his assessments of adoption and usefulness, not quantified results for every product in those categories.

His anti-use cases are more personal: stop repeating flight-booking agent demos. He wants to choose his own flights and place his own Instacart orders. That objection exposes a limit on indiscriminate delegation: a task can be technically automatable without being something a user wants to hand over. He also explicitly objects to astroturfing. The opportunity is to find valuable work users want agents to do, not merely another action an agent can perform.

9:029:14
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:02 · section reference included

Adoption evidence and a growth forecast

The final example is ChatGPT’s growth. Swyx brings in a newly published user milestone and describes asking deep research to research OpenAI and draw the growth chart. Reuters reported more than 400 million weekly active users in February 2025, compared with 300 million in December 2024—approximately 33% growth over roughly two months. That is a shorter interval than the three months described onstage. Swyx frames the larger trajectory as going from zero to that scale in approximately two and a half years.

He reads the chart as a history of model-driven demand: a period he characterizes as a year without growth, followed by acceleration when agentic models arrived. Swyx credits the o1 models with doubling ChatGPT usage. That is his causal interpretation of the adoption curve; the reported user counts do not establish which releases caused the increase.

Extending the curve, swyx predicts one billion ChatGPT weekly active users by the end of 2025, five times the September 2024 level. The slide makes that extrapolation explicit with a December 2025 billion-user callout alongside model-release annotations. It is a forecast made in the talk, not an observed milestone.

Slide titled “1B users by 2026 from Reasoning/Agents” shows a rising blue curve, model-release annotations, a “you are here” marker, and a “1B ChatGPT Users Dec 2025” callout.Open full source frame
A projected path to one billion ChatGPT weekly active users, annotated with model releases.

The product thesis follows from that interpretation: stronger reasoning and more agents shipped to users can drive demand for AI products. Swyx frames the projected billion users as roughly one-eighth of the world’s population and sees substantial commercial room for other builders. The forecast supplies the scale of his ambition; the engineering question is which capabilities will become useful products.

9:389:45
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:38 · section reference included

The AI engineer’s evolving deliverable

Machine learning engineers build models; software engineers build software; AI engineers increasingly build agents. This is swyx’s proposed direction for the profession, not an exclusive boundary between jobs. It gives the emerging discipline a concrete deliverable: systems that put model capabilities to work on tasks users are willing to delegate.

10:56
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:56 · section reference included

Resources

From the talk

Updates since the talk

  • Agent EngineeringArticle

    Swyx’s expanded companion to the keynote, covering agent definitions, enabling trends and conference examples.

Transcript download
Full transcript

Loading transcript…