The Browser Company moved from Arc to Dia after deciding that AI would change the browser itself, rather than simply improve existing browser features.
2
The team puts prompts, tools, context, models, and parameters inside its product so people across the company can prototype and refine AI features with their own context.
3
Browser security has to combine technical controls with product design, such as asking users to confirm data before Dia writes it into forms or sends actions externally.
Summary
Samir Mody describes how The Browser Company moved from Arc, which shipped in 2022, to Dia, an AI-native browser built from the ground up. The shift changed the company's product process as well as its software. Mody explains how internal tools let engineers, designers, product managers, and other employees prototype with real prompts, models, tools, and context. The team uses evals and an automated prompt-optimization process called Jeba to refine ideas before shipping. He also argues that model behavior needs its own design and engineering discipline, covering personality, instructions, context, model choice, and evaluation. Security becomes part of product design because a browser can access private data, read untrusted content, and take external actions. Technical prompt-injection defenses are incomplete, so Dia adds user confirmation before sensitive actions such as autofill, scheduling, and email writing. The move to AI required changes in hiring, training, communication, and collaboration across the company.
Arc was an improvement, but the company decided the browser itself had to change
The Browser Company began in 2019 with a mission to rethink how people use the internet. Arc shipped in 2022 and made browsing more personal, organized, and carefully designed. Mody says Arc was loved by many people, including millions of users, but the team still saw it as an incremental improvement over existing browsers. Access to GPT models in 2022 led to prototypes and some Arc features. By early 2024, the company had formed a stronger thesis: AI would transform how people use the internet and would therefore change the browser. Dia was built from the ground up around AI speed and security.
AI product teams need tools for prototyping, evaluation, data, and automation
Mody says the company wants to build tools, processes, platforms, and habits that let it iterate faster. Its AI investments include prototyping tools, evals, data collection for training and evaluation, and automation for hill climbing. The first internal tool was a basic prompt editor available only in development builds. It gave engineers limited access, slowed iteration, and lacked personal context. The company then moved prompts, tools, context, models, and parameters into the product employees use every day. That lets people from the CEO to the newest hire create or refine Dia features with full context.
Putting AI tools inside the product widens who can shape the product
The internal tools are used for more than prompt editing. The company has tools for its memory knowledge graph and for its computer-use mechanism. Mody says the team tried tens of computer-use strategies before choosing one to build into Dia. Making these tools available inside the product lets product managers, designers, customer service staff, and strategy and operations staff try ideas suited to their work. A strategy and operations team member later used the prompt tools over a weekend to rewrite all the company's prompts. His work unlocked a new level of capability and helped lead to the creation of the model behavior team.
Jeba improves prompts through scored tasks and reflective mutation
Mody describes Jeba as a sample-efficient way to improve a complex LLM system without reinforcement learning or other fine-tuning methods. The team seeds the system with prompts, runs them across a set of tasks, scores the results, and selects the strongest candidates. An LLM then reflects on what worked and what did not, generates new prompts, and repeats the process. Mody identifies reflective prompt mutation, the selection process, and the ability to tune text rather than model weights as important parts of the approach. The company uses its own metrics and scoring mechanisms to refine simple prompts.
The team separates broad ideation from disciplined refinement
The company first widens the number of ideas and lowers the threshold for who can build them. Employees try many ideas every week and every day, then use them internally. Ideas move forward when they solve a real problem, have clear utility, and appear capable of reaching the required quality. The next phase collects and refines evals to clarify product requirements. The team then improves the feature through code, prompts, and automated methods such as Jeba, continues internal dogfooding, and ships. Mody stresses that fast ideation matters because new AI capabilities create new possibilities each week, while shipping still requires a high-quality experience.
Model behavior needs its own product and engineering discipline
Mody defines model behavior as the work of defining, evaluating, and shipping the behavior a model should have. It turns principles into product requirements, prompts, and evals, then shapes the responses and personality of Dia's assistant. The work includes behavior design, data collection for measurement and training, evaluation, prompting, model selection, context-window design, and parameter choices. The process is iterative: the team builds, refines, creates evals, ships, gathers internal and external feedback, and feeds that feedback into the next cycle. Mody compares this with the growth of web product design from functional pages to more deliberate and complex experiences.
People outside engineering can be strong model behavior designers
At first, engineers wrote the prompts. After the company built prompt tools, more employees could experiment directly. A strategy and operations employee used those tools to rewrite the company's prompts and documented the work in a Loom video. The resulting prompts improved the product's capability, quality, and experience enough to contribute to forming a dedicated model behavior team. Mody says companies should look across roles for people who can shape and steer a model. The right contributors may be engineers, or they may come from strategy and operations or another part of the organization.
Browser security depends on product design as well as model defenses
Mody focuses on prompt injection, where third-party content overrides an LLM's instructions and causes harm such as data exfiltration or malicious commands. A browser faces what he calls a lethal trifecta: access to private data, exposure to untrusted content, and the ability to communicate externally by opening sites, sending email, or scheduling events. Wrapping untrusted content in tags is easy to escape. Separating instructions from third-party data with system and user roles, plus randomly generated tags, can help but gives no guarantee. Dia therefore combines technical approaches with user experience. For autofill, it shows the data in plain text and asks the user to confirm before writing it into a form.
Moving to Dia changed the whole company, not only the browser
The company first approached AI as a way to make Arc and the browser better. Mody says it soon learned that the change involved the company itself. The shift affected the creation of a model behavior team and the way the company thinks about AI security, along with training, hiring, communication, and collaboration. His final lesson is that recognizing a technology shift requires embracing it with conviction. Dia's development therefore changed the team's working methods and organizational structure alongside its product architecture and design.
"We have to blend technology approaches and user experience and design into a cohesive story that actually builds them from the ground up and solves it together."14:53
Who should watch
You are building an AI feature and need a process that lets the whole team prototype with real context before engineering turns an idea into a shipped feature.
Your product uses models to read private data and take actions such as sending email, filling forms, or scheduling events, and you need to design around prompt injection.
Your company is deciding how product design, engineering, evaluation, and model behavior work should fit together as AI changes the product.
