For most enterprise engineering leaders, the current AI conversation is centered on efficiency—how to do more with fewer people, or how to shave a few percentage points off a sprint cycle. But for Jon Hyman, the co-founder and CTO of Braze, the goal isn’t to shrink the team. It is to expand the horizon of what a 300-person engineering organization can actually build.
Braze, a customer engagement platform that has scaled from a mobile-first startup to a global leader over the last 15 years, has undergone a rapid metamorphosis. In just a few months, the company transitioned into an AI-first engineering culture. The results are stark: over 60% of the code committed to Braze’s main repositories is now AI-generated.
This shift wasn’t the result of a top-down mandate. Instead, Hyman describes a process of “enablement and guidance,” where the adoption of AI tools like Cursor, GitHub Copilot, and Claude Code was driven by a measurable increase in model quality. The tipping point came when a small team used AI to build a Model Context Protocol (MCP) server, shipping the project six weeks ahead of schedule. For Hyman, that was the “click” moment—the realization that AI had moved from a helpful autocomplete tool to a senior-level collaborator capable of driving stepwise increases in velocity.
Beyond the ‘Vibe-Coding’ Trap
Despite the surge in productivity, Hyman is cautious about a growing trend he calls “vibe-coding”—the practice of prompting AI to generate functional code without a deep architectural understanding of the underlying system. While vibe-coding can produce a working prototype in minutes, Hyman argues it is a recipe for disaster when attempting to scale.
The challenge, he notes, is that even the most advanced models cannot fit the entirety of a company’s domain knowledge into a single context window. A model might understand the syntax of a React component, but it doesn’t inherently understand Braze’s specific business processes, customer use cases, or the complex interplay of its global infrastructure. To scale, engineers must remain “on-the-ground generals,” blending AI speed with human architectural oversight.
This tension between speed and stability has led Braze to focus on codifying its “ways of working.” The company is currently transcribing its best practices—such as frontend React standards and specific testing frameworks—into a format that AI agents can use. By replacing “spaghetti skills” with standardized agentic infrastructure, Braze aims to ensure that AI-generated code adheres to strict internal patterns rather than relying on generic, off-the-shelf anti-patterns.
The Budget Shock of AI Inference
The transition to an agentic workflow has not come without a significant financial price tag. Hyman candidly describes the “budget shock” associated with high-scale AI inference. As engineers move from using AI for occasional queries to working with agents for six to eight hours a day, the cost of tokens has become a primary engineering metric.
In one instance, Hyman noted an engineer spending $150 in a single day on inference. When scaled across a 300-person organization, these costs can escalate rapidly, forcing a new kind of optimization. The goal is no longer just about the fastest output, but the most efficient use of “inference dollars per amount of output.”
| AI Evolution Stage | Capability | Engineering Impact |
|---|---|---|
| Code Completion | Auto-completing lines of code | Minor productivity gains. reduced typing |
| Junior Engineer | Executing tasks with heavy guidance | Faster prototyping; requirement for heavy review |
| Senior Engineer | Building meaningful features with minimal direction | Massive velocity increase; 60% of committed code |
| Autonomous Agent | Responding to triggers (e.g., bug reports) automatically | Shift toward 24/7 feature building and self-healing code |
Induced Demand and the Roadmap
One of the most contrarian views Hyman holds is that AI will not lead to a permanent reduction in engineering headcount for growing technology companies. Instead, he believes AI is inducing a higher demand for software.
The logic is simple: if a team that previously could only build 20 of its 100 best ideas can now build 40, the remaining 60 ideas don’t simply vanish. In a hyper-competitive market, the increase in velocity is felt globally. As every competitor also gains a stepwise increase in productivity, the baseline for “fast” shifts upward. For Braze, this means the company is nowhere near its peak engineering headcount; rather, it is using AI to attack a roadmap that was previously impossible to execute.
This shift is also redefining the relationship between engineering and product management. Product managers and designers are now using tools like Vercel v0 and Cursor to create interactive mock-ups, allowing the building process to occasionally move ahead of the final design. Because the cost of rework has plummeted, Braze can now iterate on user interfaces in real-time based on beta customer feedback, rather than waiting for a finalized Figma spec.
The integration of OfferFit, a reinforcement learning engine acquired by Braze, has further accelerated this journey, blending a fast-moving, remote-first AI culture with Braze’s established divisional structure. By treating the ML team as a parallel division, Braze has been able to integrate advanced reinforcement learning into its customer engagement tools without disrupting its core operational flow.
Looking toward 2026, Hyman anticipates the emergence of dedicated teams focused exclusively on agentic workflows. These teams will treat AI infrastructure with the same rigor as CI/CD pipelines or container orchestration, ensuring that the “brains” of the organization are standardized, durable, and scalable.
We invite readers to share their experiences with AI-generated code in the comments. Is your organization seeing a boost in velocity, or are you encountering the “vibe-coding” trap?
