Reasoning, agents, and multimodal models are expanding what AI systems can do, while model costs keep falling and competition is increasing.
2
The strongest AI applications package domain knowledge, workflow context, model orchestration, and a thoughtful user experience instead of presenting users with a generic prompt box.
3
Execution is the main source of defensibility in AI, so builders should start with workflows they understand and ship products that earn user trust.
Summary
Sarah Guo argues that AI capabilities are improving quickly enough that application builders should plan for a multimodel world with falling inference costs. Reasoning enables sequential search and other difficult knowledge-work tasks, while agents can plan, hold goals, test hypotheses, and backtrack. Voice, video, and image systems are also becoming useful in business workflows. Guo uses Cursor as the model for building AI products: the value comes from packaging context, choosing models for different tasks, fitting the product into an existing workflow, and making the output easy and safe to review. She sees similar opportunities in legal, medical, customer service, and industries that have been slow to adopt software. Full automation is often less useful than augmentation because users tolerate less failure as latency and autonomy increase. Her investment view is direct: execution, customer understanding, and trust can protect an application company even when the underlying models are widely available.
Guo describes reasoning as a new way to scale intelligence with more compute. She sees it unlocking use cases where the system must show its work, make high-stakes decisions, or perform sequential search. These resemble everyday knowledge-work problems that require several steps rather than a single answer. She expects capability to improve across the model layer and rejects the idea that application builders should prepare for the end of the current AI expansion. Model providers are also competing more aggressively, which gives application teams more choices instead of forcing them to depend on one lab.
Agents take ownership of tasks through planning and memory
Guo defines an agent in practical terms as software that takes a set of steps, includes AI, takes ownership of a task, and holds a goal in memory. It can try different hypotheses and backtrack. Agents can be simple or sophisticated, and they may use other models or search. The important distinction is that an agent does something rather than acting like a chatbot. Through Conviction's Embed grant program, which receives thousands of applications each year, she has seen the number of agent startups rise by 50% over the last year, with some already working in real-world settings.
Falling model costs make multimodel products practical
Guo says GPT-4 pricing fell from $30 per million tokens to $2 in about 18 months, while distilled versions reached 10 cents. She expects the model market to keep competing through companies such as Google, Anthropic, OpenAI, DeepSeek, SSI, and Thinking Machines, along with open-source releases. Her advice is to plan for a multimodel system and use tools such as OpenRouter or inference platforms such as Base10. Teams should be comfortable choosing different models for different jobs rather than building around a single provider.
Code became the first major AI application because its workflow is easy to validate
Guo explains Cursor's early success through properties of software development. Code is a structured logical language with significant amounts of repetitive boilerplate. Developers can validate the result by running tests, compiling, and executing it. Research labs also invested heavily in code because they view it as important to AGI, a benchmark, and a source of training data. Her main explanation is that engineers built tools for engineers and understood the workflow closely. She presents that as the pattern for other industries: domain-aware builders should redesign workflows around AI rather than wait for AI specialists to learn every domain.
The value of an AI application comes from its workflow layer
Cursor is not just one model in a text box. Guo says different models handle tasks such as diffs, merges, and file embeddings. The product packages context, prompts models carefully, reduces repetitive work, and lets teams standardize with rules. Its familiar editor, clear visual changes, and fast response make it easier for users to accept suggestions. She says builders should collect context automatically, pull it from multiple sources, use the right model at the right time, and present results thoughtfully. In her view, prompting is a stepping stone, while a good product feels informed by the user's workflow.
Conservative industries can adopt AI quickly when the product solves a real problem
Guo calls the rapid adoption of AI in low-tech industries the AI leapfrog effect. She cites Sierra resolving 70% of customer service queries for customers such as SiriusXM and ADT. Harvey, two years in, had surpassed $70 million in annual recurring revenue and had become essential to competing in legal work. OpenEvidence helps clinicians search medical research during clinical decision-making, reaching a third of US doctors weekly while the average user uses it daily. These examples support her view that strong applications begin with a clear customer and a painful workflow, rather than a generic chatbot.
Guo says co-pilots remain underrated even as interest in agents grows. She compares the progression to Iron Man's suit: it first augments the person, then handles some tasks independently. Human tolerance for failure, hallucinations, and unreliable behavior falls as latency increases, so many domains currently benefit from systems that help people rather than replace them. Her recommendation is to build the suit first and extend it toward autonomous operation as model capability improves. This lets the product deliver value without requiring users to accept an unreliable colleague from the start.
Execution and customer trust can protect an application company
Guo calls execution the moat in AI. Cursor did not invent code completion, its models, or the general product surface, but it executed across the experience and shipped faster than competitors could copy. She contrasts this with Jasper, whose first product was a collection of prompts, a text box, and strong SEO before ChatGPT displaced that approach. Builders should make products thick, understand their users, and stay ahead through repeated improvements. Customer trust also creates an advantage because the data and context needed to improve an AI experience are often generated through product use and are not readily available to the model labs.
"AI product and AI engineering, and this is kind of the theme of my talk, is quite a bit harder than people had hoped."03:15
Who should watch
Engineers deciding whether to build an AI agent, co-pilot, or workflow product will get a concrete framework for choosing between augmentation and automation.
Founders working with foundation models should watch for Guo's advice on multimodel architecture, context packaging, orchestration, and product trust.
People entering legal, medical, customer service, or other domain software can use the examples to identify workflows where domain knowledge matters more than a generic chat interface.
