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

Episode Summary

Executive Summary: The conversation centers on Vercel’s AI strategy after Ship AI: making agentic development practical through workflows, low-level AI SDK primitives, and tightly integrated production tools. The CTO explains how Vercel grounds AI products in real internal use, open sources useful abstractions, expands into Python, and is building agents for DevOps, sales, abuse review, and data analysis while emphasizing security, human oversight, and native-fit products.

Main Topics: Vercel’s AI strategy and Ship AI recap (Priority: 5/5): The CTO frames Vercel as a strong supporter of AI engineering, focused on turning hype into concrete developer tools and products that are grounded in real usage. Workflow Development Kit and workflow abstractions (Priority: 5/5): A major announcement is a workflow development kit that makes durable, streamable, retryable, long-running workflows feel first-class for agents and apps. AI SDK v6 and the low-level approach (Priority: 5/5): AI SDK stays intentionally low-level so it can adapt as the AI app landscape evolves, with version 6 introducing a direct agent abstraction after patterns became clearer. Vercel Agent and internal agent use cases (Priority: 5/5): Vercel is building agent products for code review, DevOps, sales qualification, abuse analysis, and data analysis, leveraging first-party runtime, logs, and deployment context. Open source and business model philosophy (Priority: 4/5): The discussion explains Vercel’s open-source approach: truly open projects that grow adoption and monetize through ecosystem value rather than restrictive licensing. Python support and infrastructure expansion (Priority: 4/5): Vercel is investing heavily in Python support, including zero-config deployments for Flask/FastAPI, a Python SDK, and fluid compute for long-running AI backends. Leadership, organization, and security for AI-era app building (Priority: 5/5): The CTO discusses company structure, IC leadership, agent adoption, and a future where apps are secured even when both developer and AI are imperfect.

Key Arguments: Vercel’s AI work is most effective when it is based on products the company actually uses internally, not abstract framework design. Workflows are essential infrastructure for AI applications because agents need durable, resumable, retryable execution. AI SDK succeeded by staying low-level and avoiding premature assumptions about how AI apps should be built. Agent abstractions should emerge from real usage patterns rather than being forced in from day one. Vercel Agent can be more useful than generic coding agents because it has access to runtime data, logs, deployment metadata, and secrets. Open source can be a strong business model when the software is truly usable everywhere and the company monetizes ecosystem participation. Python support is strategically necessary for a serious AI cloud, and Vercel is investing in native-feeling support rather than superficial compatibility. The next frontier is security and infrastructure that assume both the developer and the AI may make mistakes. Agents are most effective on boring, repetitive, judgment-light tasks where the business impact is high and human time is wasted. Human oversight still matters; Vercel is not claiming agents can safely replace all operational decisions, especially high-risk changes.

Data Points: Vercel CTO tenure: a little less than four years - He joined Vercel before ChatGPT and describes transforming the company for the AI era. AI SDK version: v6 beta - Announced during the discussion, with a new direct agent abstraction becoming stable. Workflow wait duration: multiple days - Workflows can pause and resume for days without cost or issue. Compute cost during workflow wait: literally does not cost anything - Paused workflow steps do not consume billing while waiting. Internal reports: 2 - He mentions having two direct reports in the CTO organization. Company commitment for agent adoption: 3 agents - For the forward-deployed assistance program, companies are asked to commit to building three agents. Time-series example latency: 30 seconds - Used to illustrate why streaming and long-running AI operations matter more than old millisecond assumptions. Typical old latency benchmark: 500 milliseconds - He contrasts traditional software expectations with AI model latency. TypeScript ecosystem size: biggest language on GitHub - He cites this as evidence that TypeScript remains very relevant alongside Python.

Pivotal Quotes: "we never give you an abstraction that we haven't used ourselves" — CTO of Vercel: Describing Vercel’s product philosophy and how framework ideas are validated before release. "agents are both extraordinarily effective and still very ineffective" — CTO of Vercel: Explaining why Vercel focuses on bounded, practical use cases rather than broad autonomy. "I want to be able to build an app that is secure even if the developer is incompetent" — CTO of Vercel: Discussing the future of AI-era app infrastructure and security assumptions.

Implications: Vercel is positioning itself as an AI application platform, not just a deployment host. Expect more workflow tooling, agent products, Python support, and security primitives aimed at making practical AI systems shippable and trustworthy.