On July 22nd, GV hosted our first AI Builders NYC event, bringing together our portfolio founders and the community of builders in New York. I invited Zach Lloyd, founder of Warp, to join me to share his insights on the rapidly evolving landscape of AI in software development. Our conversation explored the technical underpinnings of this shift, from the models powering the next generation of coding tools to the evolving role of the engineer. Here are some of the key insights from our discussion.
When GV invested in Warp in April 2020, the vision was to reinvent the 40-year-old developer terminal. Today, that vision has expanded dramatically. Warp is no longer just a terminal; it’s what Zach calls an “agentic development environment.”
The Three Stages of AI in Coding
Zach framed the recent evolution of AI-powered coding in three distinct stages. We’ve moved from the first phase, dominated by inline completions (like those popularized by GitHub Copilot and Cursor), into a more powerful second phase.
“Today, and even just really in the last couple of months, you’re seeing the transition to a more agentic style of coding,” Zach explained. “You don’t even need the file editor to be your primary interface… Instead, what you do as a developer is you write a prompt and you specify a bunch of contexts that you think are important for the agent to solve the problem. You hit Enter, and the agent starts solving it for you.”
This “Phase Two” is where tools like Warp, Google’s Gemini CLI, and Anthropic’s Claude Code operate. The primary interaction model shifts from writing code to expressing intent in natural language.
Looking ahead, Zach predicts a third stage driven by system events. Instead of a developer manually initiating a task, an agent will be triggered automatically. How will this happen? “A crash has been detected through your crash reporting system, let’s have an agent try and fix that,” he posited. “Or a user has filed an issue about a misalignment in the UI, and it will start to fix it for you.” In this future, models will run autonomously for longer, with the developer’s role shifting to that of a conductor, providing high-level guidance and intervening only when an agent hits a wall.
A Multi-Model Strategy for a Competitive Landscape
With the explosion of options for foundation models, choosing the right one is critical. Warp’s approach is to remain model-agnostic while providing smart defaults.
“We default today to the models from Anthropic, specifically Claude Sonnet and Opus,” Zach noted. “We have a whole setup internally where we eval all the models and have found that those are at the frontier for doing coding right now.”
However, Warp is designed to give users a choice, allowing them to leverage the unique strengths of different models, like Gemini’s large context window. This flexibility extends to their internal architecture, where they use a multi-model strategy.
“Sometimes you want a very… high latency, very expensive reasoning model,” he said. “Sometimes it’s better to have a very low latency, fast model for classifying something. So we try and be intelligent here.” This nuanced approach—using the right tool for the right job—is essential for building a high-performance, cost-effective product on top of LLMs.
Winning on Product, Not Just Price
In a market where tech giants can give away model requests at a massive scale, competing on cost is a losing battle for startups. For Warp, differentiation comes from its unique position in the developer’s workflow.
“Most of the coding tools in the market are either in the IDE, specifically in a fork of VS Code… or they are CLI apps,” Zach observed. “For Warp, we have a totally different spot in the stack where we are basically this thing that kind of looks and feels like a terminal but is really an agent platform and there’s nothing else like it.”
This architectural advantage is a key reason Warp has performed exceptionally well on industry benchmarks like SWE-bench (top five) and Terminal Bench (number one). Because Warp is the terminal, it can observe the full context of a developer’s session—every command, output, and error—giving its agents a powerful advantage in completing complex tasks. The goal is to create lasting value and “developer love” by solving problems across the entire development lifecycle, from coding a new feature to debugging a production fire.
The Evolving Engineer and the Skill Gap Paradox
The rise of agentic workflows raises a crucial question: What happens to the fundamental role of an engineer? If an agent can build an application from a prompt, do developers risk losing their understanding of data structures, system architecture, and optimization?
Zach acknowledged this is a “real risk.” The current solution at Warp is a policy of ownership and explicit instruction. “As an engineer today at Warp, you are responsible for all code that gets submitted for review,” he stated. “We ask them to tell the agent not just what they want built, but how to build it… specify what data structure you want, what you want the API to look like.”
This is a temporary bridge. In the long term, our entire software development ecosystem needs to evolve. We’ve seen this before; developers moved from managing memory and registers to writing in high-level languages, trusting the layers of abstraction beneath them. We are at a similar inflection point.
The future won’t be about humans reviewing every line of machine-generated code. Instead, we’ll need new infrastructure and interfaces for validation, security scanning, and performance testing—perhaps, as Zach mentioned, a new kind of code review interface designed for human-to-agent collaboration. The end goal, Zach predicts for 2030, is not the elimination of engineers, but the creation of “a lot more software,” built by creative problem-solvers who have been freed to operate at a higher level of abstraction.