Podcast Summary

Podcast: Latent Space: The AI Engineer Podcast
Episode: ⚡️ Ship AI recap: Agents, Workflows, and Python — w/ Vercel CTO Malte Ubl
Host: Latent.Space
Guest: Malte Ubl, CTO of Vercel
Date: October 31, 2025
Theme: A deep dive into Vercel’s innovations from the Ship AI conference, focusing on AI agents, workflows, open-source frameworks, Python support, and the evolving role of AI engineers.

Episode Overview

This episode recaps Vercel's Ship AI event, with CTO Malte Ubl offering a comprehensive look at Vercel’s latest moves in AI engineering—particularly around AI agents, workflow abstractions, product strategy, and their approach to open source and Python. The discussion explores how Vercel builds for AI engineers, the practical realities of durable workflows, agent-powered developer tools, the challenges and philosophies of framework design, and lessons learned on leadership and organizational transformation in the AI era.

Key Discussion Points & Insights

1. Vercel’s Vision for AI Engineering

[01:16] – [03:19]

• Malte Ubl outlines Vercel’s commitment to AI engineering, not just by riding hype, but by building concrete abstractions and products ("We’re the biggest fan of the AI engineering movement... being very concrete about, like, agents are very exciting and you can actually build them." – Malte, 01:16).
• The Ship AI conference centered on making agent and workflow development easier with durable, composable abstractions—anchored in building real apps as a way to ground product decisions.

2. The Rise and Rationale of Durable Workflows

[03:19] – [08:06]

• The conversation delves into why workflows matter and how they're underappreciated and under-taught, despite their ubiquity in production systems.
  • "You don't really learn about this in CS classes. ... It's not really a unit of compute and storage that is taught—it's emergent from Uber and Stripe and everyone else." – Host, 03:19
  • Malte details the historical roots and how every serious transactional system (from '70s banks to modern tech giants) reinvents some form of workflow abstraction (“…there either is an explicit abstraction for workflows or someone made one ad hoc because otherwise the thing just doesn’t work.” – Malte, 04:48)
• Vercel’s new Workflow Development Kit lets developers express long-running, resumable tasks in idiomatic code, handle human-in-the-loop steps with webhooks, and do this efficiently, open-source and cost-free.
  • "You can run compute for infinite amount of time... Automatically retry them, stuff like that. That makes things more reliable." – Malte, 05:17

3. Philosophy of Open Source and Product Design

[06:47] – [08:01]

• Malte positions Vercel's open-source strategy as maximizing the collective benefit and reach: "Our strategy is to grow the pie while what we actually see is that our pie piece is a relatively constant size in proportion to the pie... it’s a business model that has more winners than the alternatives." (06:47)
• Production Considerations: Even when most users won’t self-host, open-source code gives teams a sense of control, auditability, and optionality (“you want to check the checkbox 100%... I actually do think people will run it themselves and that's great.” – Malte, 08:01).

4. Patterns & Principles Behind Vercel’s AI SDK and Frameworks

[09:26] – [17:47]

• AI SDK’s Success: Attributed to humility in abstraction—keeping things low-level and letting patterns emerge from real usage rather than assuming what users want.
  • "If you put a very thick abstraction, then it's probably going to be the wrong abstraction. So you have to be humble..." – Malte, 11:44
• Contrast with Big Labs: While some AI labs push for high-level, “let-the-model-drive-everything” architectures, Vercel prefers practical, human-first tooling: "We are at the jQuery era. We don’t know what we want yet. We're building all these tools to make the smallest possible things easier." – Host, 13:22
• Dogfooding Principle: Vercel never ships an abstraction they haven’t used themselves, ensuring their frameworks are battle-tested (“Dogfooding is ultimately the thing. ...There's this constant feedback loop where if you don't have that...framework builders are usually not application builders.” – Malte, 17:47).

5. Agent Use Cases — Product and Internal

[18:33] – [29:44]

• Distinguishing Agent Types: Vercel has agents as product (Agent-as-a-Service, e.g., Vercel Agent for DevOps and observability) and custom internal agents for their own ops/sales/support (“We distinguish between agent as a service and stuff we run internally. They are not the same thing.” – Malte, 18:49).
• DevOps/Observability Agent: Anomaly detection triggers the agent, which uses observability queries and log analysis to diagnose issues, acting as a "coworker that has no sleeping problems in the loop" (21:57). This enables aggressive anomaly detection and smarter escalation—“It’s just so much easier than doing it yourself.”
• Where Agents Shine & Don’t:
  • Agents excel at tedious, repetitive, judgment-light tasks (“Ask your company: what do you hate most about your job? ...These problems are probably easy enough for a current generation agent to handle.” – Malte, 26:15)
  • Use cases open-sourced: sales lead qualification, abuse analysis (pre-work for human review), and a structured data analyst agent.
  • Agents aren’t yet safe for high-risk actions (e.g., firewall changes, DNS migrations: "I would not let the agent do that yet. Right. It's just too dangerous." – Malte, 25:47)

6. Forward-Deployed AI Engineering & Customer Empowerment

[29:55] – [31:41]

• Agent on Every Desk: Vercel’s program offers direct engineering support to help customers build their first AI agents—accelerating adoption and collecting valuable product insights.
  • “As a startup, I just want to see the open source project...as a large company, it’s daunting to ship the first agent; so something like a forward deployed engineer does help.” – Malte, 29:55

7. Python Support, Fluid Compute, and Language Agnosticism

[32:13] – [35:26]

• Expanding Beyond TypeScript: Vercel now supports Python zero-config for prominent frameworks (Flask, FastAPI), released a Python SDK, and is making investments to ensure parity in experience ("It is also on a Fluid Compute program... you only pay when you have compute." – Malte, 33:25).
• No Language Wars: Vercel doesn’t take sides, but recognizes supporting both Python and JavaScript/TypeScript is table stakes for serious AI platforms (“Honestly, obviously I don't really care. I think both communities are very relevant, very large, and we are investing in supporting them.” – Malte, 34:27).

8. Leadership, Org Change, and Secure Agent-Native Infra

[35:26] – [41:05]

• CTO Reflections: Malte shares lessons on evolving Vercel through the AI revolution, stressing the importance of "playing to your company's strengths" and building native-feeling AI products (“You have to be honest with yourself. What product...does kind of extend naturally from what I’m doing, rather than ...something that’s entirely different.” – Malte, 36:04)
• IC vs. Management Track: Inspired by Google’s promotion ladder, Vercel keeps strong contributors on the IC path rather than forcing management (“You have to be willing to live in this world where you don't make your strongest engineers have the choice of not making more money or becoming a potentially very bad manager.” – Malte, 38:19)
• AI for All Builders, Not Just Developers: Designing agent-native infra that’s secure “even if the developer is incompetent” and separates sensitive logic (auth, data-level access) from the AI-generated/modified apps—preparing for a world of code built by designers, PMs and agents (“Assume the developer doesn't know what they're doing and ...they’re using an AI that also doesn't know what they're doing. ...A way to build an app that is secure even if the developer is incompetent.” – Malte, 39:32)

Notable Quotes

• On practical AI progress:
  “Agents are both extraordinarily effective and still very ineffective. ...You have to find the right problems. And then when you find the right problems, they are super magical..."
  – Malte Ubl, 25:57

• On workflows and abstraction:
  "When I run a bank transaction system in 1975, then I invent this, right? ...There's a version of [workflows] at every single company that has been doing anything in computer since 1950."
  – Malte Ubl, 03:46

• On humility in product design:
  "We know absolutely nothing and we still know absolutely nothing. ...If you put a very thick abstraction, then it's probably going to be the wrong abstraction."
  – Malte Ubl, 11:44

• On open sourcing critical infra:
  "Our strategy is to grow the pie...it’s a business model that has more winners than the alternatives."
  – Malte Ubl, 06:47

• On AI agent limitations:
  "I would not let the agent do that yet....It's just too dangerous."
  – Malte Ubl, 25:47

• On secure app platforms for the AI era:
  "We are very deeply working on a way to build apps that follows the threat model that assumes the developer doesn't know what they're doing and also they're using an AI that also doesn't know what they're doing."
  – Malte Ubl, 39:32

Timestamps for Important Segments

| Timestamp | Topic | |-------------|------------------------------------------------| | 01:16 | Vision for Vercel & AI Engineering | | 03:19 | Hidden history and import of workflow systems | | 05:17 | Infinite/lazy compute in Vercel workflows | | 06:47 | Vercel's open source/product philosophy | | 09:26 | Why Vercel’s AI SDK succeeded | | 13:22 | Agent abstraction differences (labs vs frameworks) | | 18:49 | Agent as service vs Internal agent use-cases | | 21:57 | DevOps/Observability Agent explained | | 25:57 | Agents: “Extraordinarily effective, still ineffective” | | 26:15 | Where agents excel (tedious/boring tasks) | | 29:55 | Agent On Every Desk: Forward deployed engineering | | 32:13 | Python zero-config & SDK, Fluid Compute | | 35:26 | CTO/leadership evolution in AI orgs | | 39:32 | Agent-native, secure-by-default app design | | 41:43 | Wrap-up and closing reflections |

Memorable Moments

• Dogfooding frameworks: Vercel insists on building real products with their own abstractions, ensuring frameworks aren’t “ivory towers.” (17:47)
• Practical AI agent boundaries: Malte recounts giving agents aggressive anomaly detection because “they don’t sleep,” but draws the line at automating DNS migrations (“That would be AGI.” – Host, 25:50)
• Embracing AI coding for all: The future is apps (securely) “vibe-coded” by designers, PMs, and agents—not just traditional engineers. (39:32)

Conclusion

This discussion provides a masterclass in pragmatic AI engineering, open-source strategy, and product-driven abstractions. Malte Ubl’s insights illuminate the messy, emergent, and exciting path from “hype” to real agent-powered workflows—grounded in real usage—and lay out how Vercel is evolving to empower all builders in the era of AI.

For more resources, open source code, and detailed show notes, visit latent.space.