⚡️ Ship AI recap: Agents, Workflows, and Python — w/ Vercel CTO Malte Ubl

Our high-level goal is to champion the AI engineering movement. We're not just focused on hype and big ideas; we aim to be concrete. For instance, agents are exciting, and you can actually build them. Our conference focused on simplifying this process, discovering the right abstractions as user needs emerge. Vercel approaches this by building things ourselves. This includes both products, such as agents you can purchase from Vercel, and internal tools to enhance our own operational efficiency.

Every company involved with computers since 1950 has faced a version of this challenge. However, it hasn't always been abstracted. For instance, if I ran a bank transaction system in 1975, I would have invented a solution. Even in a pure batch processing world, I would have built something similar. So, I either use a productized solution, which Temporal innovated, or I create an ad hoc system. This might involve a queue to track work, a database to store state, and perhaps a cron job to prevent items from getting stuck in the queue.

I view open source as having three business models. First, there's Red Hat, where you self-support open-source software. Second, there's open core, where you're the sole monetizer, but others can still run it. Vercel, though perhaps not inventing it, has certainly popularized a model where a truly open-source software library, with adapters for every platform, becomes widely popular. We then get a piece of that pie. Our strategy is to grow the overall pie, as our share remains a relatively constant proportion. This allows us to drive the open-source project.

I'm not sure how much that helps, but it's great. There's a fun background from what people thought AI was ten years ago. We announced version 6 beta, and while it's not news since these things are open source, it introduces a stable feature: direct agent abstraction. This wasn't previously available as a stable feature, though it was experimental in ASDK5. People would build agents with AISDK, but they had to do it in a more complex fashion. I want to mention that it's very successful, and I think the reason is that we constrained ourselves to be humble about what our users might want to do.

I want to be in control, configure settings, and define defaults as I see fit. It's good to have tension around these things. A core principle at Vercel, established by Guillermo, is that we never provide an abstraction we haven't used ourselves. Dogfooding is crucial. The AI SDK was extracted from V0. We built it, diverged slightly, then undertook significant work to fully host AI SDK within V0. This process allowed us to learn and ensure migrations weren't too difficult, which users appreciate. This constant feedback loop is essential. While it sounds obvious, framework builders are often not application builders.

From a technical standpoint, this agent has several tools. It can execute any observability query against your project using its query builder. It can also read logs, often with queries. What's truly magical is its proficiency; when you click on an anomaly, it almost always precisely tells you what happened. It displays all the relevant graphs. It's significantly easier than doing it yourself, saving minutes of work. What excites me most, and this is a common pattern with agents, is that in many situations, it addresses the recall precision problem.

While beneficial, this approach isn't suitable for every company. For large organizations, there's significant efficiency to gain, but deploying a first agent can be daunting. In such cases, a forward-deployed engineer, which we provide, is very helpful. However, as a startup, I wouldn't want a forward-deployed engineer in my office. I'd prefer to use an open-source project, integrate it with cloud code, and then apply it to my specific problem, customizing it as needed. This should also be successful. Our program aims to unblock individuals who are overwhelmed by the hype and unsure how to select the right project.

I joined Vercel a little less than three or four years ago, before ChatGPT. I came directly from Google, so I had some insights into what was happening. However, Vercel wasn't operating in that world; it was the tail end of the last big crypto wave. Then, the AI revolution suddenly occurred. I definitely had to work through how to transform the company into something quite different. The solution we've found feels really good, and there's a lesson for companies that haven't made this transition yet: you must do something that feels native to your company.