Episode

The Post-Coding Era: What Happens When AI Writes the System?

Nicholas Moy, former Head of Research at Windsurf & now at Google DeepMind, joins High Signal to discuss the shift from "co-driving" to a truly "agentic" era of development. We discuss Windsurf's journey from early prototypes that struggled with compounding errors to the successful launch of their agentic coding product. Nick explains that building a startup in the current climate requires a strategy of "disrupting yourself" to avoid the innovator’s dilemma; companies must be ready to pivot as soon as a new frontier model makes previously impossible features viable. He argues that traditional technical moats are increasingly fragile, and true defensibility now comes from real-world usage data, brand reputation, and a deep intuition for what users need at the frontier of these capabilities.

January 12, 2026

Listen on

Guest

Nicholas Moy

Google DeepMind

Nicholas Moy was the Head of Research at Windsurf where he built the first multi-step coding agent. He now works with the Antigravity team at Google Deep Mind, focusing on unlocking the agentic reasoning capabilities of Gemini models.

Guest

HOST

Hugo Bowne-Anderson

Delphina

Hugo Bowne-Anderson is an independent data and AI consultant with extensive experience in the tech industry. He is the host of the industry podcast Vanishing Gradients, a podcast exploring developments in data science and AI. Previously, Hugo served as Head of Developer Relations at Outerbounds and held roles at Coiled and DataCamp, where his work in data science education reached over 3 million learners. He has taught at Yale University, Cold Spring Harbor Laboratory, and conferences like SciPy and PyCon, and is a passionate advocate for democratizing data skills and open-source tools.

‍

Key Quotes

Key Takeaways

Engineer for Problem Invariants

In an era where models double in capability every few months, Nick suggests building software around “information and output invariants”: the fundamental data requirements and final work formats of a task. By focusing on what must be true regardless of the AI's power, teams can avoid the trap of re-writing their entire core logic every time a new frontier model is released.

‍

The "Double Penalty" of Agent Errors

Shifting from co-driving (real-time assistance) to agentic engineering introduces a new risk: the "double penalty." If an agent fails, the human must pay the cognitive cost to "download" and understand the agent’s complex state before they can even begin to fix the error, making agent self-explanation and verification artifacts more valuable than the code itself.

‍

PRs are Replacing the IDE

As agents move toward true autonomy, the editor becomes a secondary interface. The primary surface for supervising AI is shifting toward "agent manager modes" that resemble Pull Requests, where users review high-level artifacts, screen recordings of UI tests, and unit test logs rather than watching a cursor move in real-time.

‍

Synthetic Data is not the Answer

While synthetic data offers scale and cleanliness, real-world user data is the only way to capture "soft, fuzzy distributions" of human behavior. Nick notes that the most durable data moat is learning "negative" preferences, such as the subtle, intuitive moments when a model should not show a suggestion, which is nearly impossible to generate synthetically.

‍

Personalization is the Application Moat

While foundation model labs dominate general intelligence, application-specific startups have a "prime position" to win on per-user model personalization. By owning the direct relationship with the user, these companies can build feedback loops that tailor model behavior to individual styles in a way that centralized labs cannot replicate.

‍

Be Prepared to Disrupt Your Cash Cow

Windsurf’s success came from a willingness to "prepare for disruption" while their autocomplete business was still thriving. They kept an agentic framework on the back burner for 18 months, periodically "testing the harness" with new model checkpoints until Claude finally "clicked," allowing them to pivot the product in just one month.

‍

The Strategy of Betting On Capabilities

With the pace of model improvement accelerating, Nick encourages teams to bet on future capabilities. If a feature seems impossible or too expensive today, building the infrastructure for it now (as if the model were already 50% more capable) allows you to capture the market the moment the next frontier checkpoint drops.

‍

You can read the full transcript here.