Browse · People
Person · Latent Space wiki

Malte Ubl

Software engineer and Vercel CTO in 2025–26 whose public work addresses agent reliability, evaluation and security, following a career building web frameworks at Google.

Malte Ubl is a software engineer whose work spans web frameworks and AI infrastructure. As Vercel’s CTO in 2025–26, he advocated narrowly scoped agents, ordinary code for deterministic tasks, and security boundaries that keep generated programs away from credentials. He also coauthored accounts of v0’s evaluation practices and announced Vercel Agent’s public beta in June 2026.[3][4][5][6][7]

Building agents with ordinary software

In his June 2025 guide to agent development, Ubl argued that effective agents should be narrow and domain-specific. His starting point was practical: “Start by simulating the agent manually, before writing any code.” Developers should use real inputs, guide the model through the task, and carry out its proposed actions themselves before deciding whether to automate the process.[3]

Ubl illustrated the value of scope with the v0 team’s experience in 2023. He recalled that models initially struggled to generate reliable HTML; the breakthrough came from narrowing the task and using structured inputs with Tailwind, rather than from better models. Once a task showed promise, he recommended implementing a loop or simple state machine: collect inputs, perform deterministic computation, call the model where judgment is needed, and decide whether to continue.[3]

Evaluation was another documented contribution. In an October 2024 article coauthored with Alice Alexandra Moore and Ido Pesok, Ubl described evals as checks of output quality that complement an existing test suite. Their account of v0 combined code checks, user and internal human feedback, and LLM grading. An automated suite reported regressions and improvements, Braintrust retained logs for manual review, and every pull request affecting the output pipeline included eval results. The authors also acknowledged that maintaining the suite required ongoing updates.[6]

In the October 2025 Ship AI conversation, Ubl discussed building abstractions from internal production use. Latent Space’s account connected this dogfooding philosophy to extracting the AI SDK from v0. He also discussed durable workflows that let code pause and resume, including human approvals through webhook patterns.[1]

Containing agent failures

Ubl’s June 2025 security guidance treated prompt injection as a design constraint. He recommended assuming that an attacker could control the model’s prompt and limiting tools to the caller’s authority. His tenant-analytics example made the boundary concrete: letting the model choose a tenant ID could expose another tenant’s data; binding that identity when creating the tool restricted the model to the correct tenant.[4]

He also explained how output could become an exfiltration channel. A model-generated Markdown image URL could contain sensitive data; rendering the image would cause the browser to send that data in a request, even if the model could not make network requests directly. Ubl recommended sanitizing output before rendering or passing it to other systems. His governing principle was: “Security is not about trusting the model. It is about minimizing damage when it behaves incorrectly.”[4]

In February 2026, Ubl and Harpreet Arora developed this position in Security boundaries in agentic architectures. They recommended running the trusted agent harness and its credentials separately from generated-code execution. A sandbox shared by both could protect the surrounding infrastructure while still leaving the harness’s credentials accessible to generated programs. Their preferred architecture combined isolated execution with network-level secret injection, so generated code could use authorized services without reading raw credentials. They explicitly distinguished preventing secret theft from preventing misuse: a secret-injection proxy alone could still permit unwanted API calls during runtime.[5]

At Ship 2026, Ubl announced Vercel Agent’s public beta, along with Dockerfile support and Vercel Container Registry. Vercel’s June recap described the agent as investigating production anomalies and proposing fixes through pull requests for review. Its stated permissions model used a separate identity, read-only defaults, and narrow, temporary permission requests before touching production.[7]

Before this AI work, Ubl built web frameworks at Google and helped organize the JavaScript community. That earlier career supplies context for his later writing about frameworks and software development; by June 2022, he had left Google and joined Vercel.[2]

Earlier career: Google and web performance

Ubl recalled joining Google in Hamburg in 2010. Gergely Orosz’s June 2022 career interview described eleven years there, progressing from Software Engineer III to Principal Engineer. Ubl built Wiz, a framework used by Search, Gmail, Photos, Drive, Meet and other consumer interfaces. Orosz also credited him with a central role in creating AMP and applying its lessons to the Page Experience ranking system based on Core Web Vitals.[2]

On October 1, 2021, Ubl announced that Google Search had ended Internet Explorer 11 support in its main product. Search remained available through a fallback experience. His public message encouraged developers to use the announcement when making the case to their own employers to move beyond the browser.[2]

Early employment, JSConf and management

Ubl’s programming experience began earlier still. He recalled learning Smalltalk in an after-school software job at fourteen in the mid-1990s, and valued its interactive feedback. At eighteen, while still in high school, he joined Schaffhausen ITC Solutions. The company paid his private-university tuition in exchange for an expectation that he would stay after graduating. He later managed its web unit and credited participation in sales with teaching him skills beyond engineering.[2]

After moving to SinnerSchrader in 2009 to do more engineering, Ubl attended the first JSConf in Washington, DC that spring. He and another German attendee decided to organize a Berlin conference, which took place that November. He recalled remaining involved through the last event in 2019, learning both fixed-deadline delivery and the opportunity to rethink an event each year. Vercel’s 2026 recap places his first meeting with Guillermo Rauch at JSConf EU at Alte Münze in Berlin.[2][7]

Ubl described his movement between management and individual-contributor roles as an oscillation, with management more a necessary responsibility than a passion. In 2022, he argued that becoming a manager should be a sideways move rather than bundled with promotion, and that engineers should be able to return to individual-contributor work if management did not suit them.[2]

From the Latent Space archive

Original posts and separately reviewed explanations.

Sources

  1. 1
  2. 2
  3. 3
  4. 4
    Building secure AI agentsAccessed 2026-09-18
  5. 5
  6. 6
  7. 7
    Vercel Ship 2026 recapAccessed 2026-09-18

Model source and teaching reviews passed. Human editorial review pending. This profile describes the sources as of 2026-09-17.

Latent Space

Join Latent Space.

Sign in with Google, or get the newsletter by email.

Loading Google sign-in…

or subscribe by email

Free. Unsubscribe any time.