# Treasure AI Night Recap — Behind the Scenes of AI-Native Products: From Research to Enterprise Product Toru Takahashi Note This is a recap of an in-person event held in Japan. The talks were delivered in Japanese. Event page: [Treasure AI Night on connpass](https://treasure-data.connpass.com/event/391551/). On June 2, 2026, Treasure AI hosted **Treasure AI Night — Behind the Scenes of AI-Native Products**, an event for engineers and product builders. In April 2026, we rebranded from Treasure Data to **Treasure AI**. The name change is more than a refresh of our brand. It signals a new direction: rather than offering AI as just one more feature, we are redesigning the company and our products themselves around the assumption that AI is a given. For this Treasure AI Night, our theme was **"How does an AI-native company build products?"** We shared the AI development platform we use internally every day, as well as the architecture of the enterprise AI platform we have newly developed. The event was held at the office of **Findy Inc.**, where many engineers and product developers joined us for an active discussion about AI agents, AI-native development processes, and enterprise AI infrastructure. ## Rebranding to Treasure AI ### Redefining the company in the age of AI In the opening session, Product Manager Toru Takahashi introduced the rebrand to Treasure AI and our future product strategy. Treasure Data was founded in the United States in 2011, and has evolved from a big data platform into a Customer Data Platform (CDP). But the spread of generative AI has dramatically changed the environment surrounding enterprises. AI models themselves are rapidly becoming commoditized. Going forward, competitive advantage will come not from *which model you use*, but from **what data, business processes, and permission controls you can connect to AI**. With this shift in mind, Treasure AI has set out a new vision: the **Agentic Experience Platform**. To realize AI agents that understand enterprise data, assess situations, and act autonomously, we are building a platform that integrates four layers: - **Perception** - **Memory** - **Intelligence** - **Action** To deliver on this vision, we are developing a family of products: - Treasure Work - Treasure AI Studio - Engage Studio - Treasure AI Voice - Composable Audience Studio ## Treasure Work ### Building an environment where every employee can use AI agents Next, Senior Principal Software Engineer Taro L. Saito spoke about the background behind **Treasure Work**, our internal AI platform. When it comes to adopting AI, attention tends to gravitate first toward the latest models and agents. But the challenge we faced was somewhere else entirely. We rolled out Claude Code company-wide in 2024, yet six months later the adoption rate was stuck at only around 30%. The reason was simple. Developer-oriented interfaces — terminal operations, Git, Unix commands — were a high barrier for many employees. That is why we built Treasure Work. Treasure Work is a desktop application based on Claude Code that pre-packages everything needed to use an AI agent. Users can start using an AI agent simply by installing it, with no complex setup. Today it has become an internal standard tool used by **more than 80% of employees**, across a wide range of work: - Software development - Documentation - Research - Data analysis - Product planning ### The Three-Layer Harness Through developing Treasure Work, we arrived at a design philosophy we call the **Three-Layer Harness**. An AI agent's capability is not determined by model performance alone. What tools it can use, what knowledge it is given, and what UI it is used through — all of these matter. So we separated our AI platform into three layers: - **Tool Layer** — the execution foundation an agent acts through: CLI, MCP, the file system, and various APIs. - **Knowledge Layer** — the knowledge an agent uses to make decisions: skills, guidelines, system prompts, and best practices. - **UI Layer** — the interface a user interacts with the agent through: CLI, desktop app, and web app. This structure lets us reuse our assets even as models and agents change, and lets us bring research results into products quickly. We also adopted a staged structure: - **tdx** — a CLI-based research environment - **Treasure Work** — a GUI-based experimentation environment - **Customer-facing products** This creates a cycle in which we can validate new technology safely and quickly, then fold it into our products. iframe [Download slides (PDF)](/assets/treasure-ai-night-2026-three-layer-harness.8bd6a34c5f3d86204db952293b8b1fff638d43f55e85ab989012f4aa7cc8d5d0.467246ce.pdf) ## Treasure AI Studio ### The execution platform behind enterprise AI Co-Founder and Chief Architect Sadayuki Furuhashi introduced the architecture of our new product, **Treasure AI Studio**. Treasure AI Studio is an AI-native platform designed for enterprises. In this session, he shared both the development process from planning to release, and the technical foundation behind it. What stood out most was the development speed. The first plan file was created at the end of February 2026. About 10 days later, a working prototype was complete. After adding support for web, mobile, and desktop, the product reached general availability in roughly three months. This speed was made possible by an AI-native development style centered on Claude Code. By collaborating with AI through design documentation, architecture reviews, implementation, and pull request creation, we drastically reduced the waiting time and communication costs that the traditional development process incurred. ### An "AI OS" for running AI agents safely At the same time, the more powerful the permissions you grant an AI agent, the larger the security and governance challenges become. AI agents operate while accessing: - the file system - the network - code execution environments - external APIs What is convenient for individual use requires careful control in enterprise use. To solve this, Treasure AI Studio includes a purpose-built **AI OS (AiOS)**. By combining technologies such as Bottlerocket OS, MicroVM, Linux namespaces, seccomp, OverlayFS, and SSL inspection, it achieves multi-layered security. It also incorporates mechanisms built for real-world operation: - pre-launching of sandbox environments - low latency via an execution-environment pool - automatic recovery when an agent fails - audit log collection Rather than being a mere chat tool, Treasure AI Studio is designed as an AI execution platform that enterprises can use with confidence — and that is its defining characteristic. iframe [Download slides (PDF)](/assets/treasure-ai-night-2026-treasure-ai-studio.b6f38bdfa2c296473c7cde3f253d4b074015e749c45bb2ba8657e1e936536648.467246ce.pdf) ## What we aim for as an AI-native company What we wanted to convey at this Treasure AI Night was not "adopting AI" in itself. What matters is **redesigning the company, the products, and the development process themselves around the assumption of AI**. The insights gained from Treasure Work flow back into our products. The lessons learned from product development are fed back into research and development. We believe that keeping research, internal use, and customer-facing products in a continuous loop — never siloed — is exactly what makes an AI-native company competitive. Treasure AI will continue to actively share its practical knowledge of AI agents, AI-native development processes, and enterprise AI. ## Closing Thank you to everyone who joined us, and to **Findy Inc.** for providing the venue. Treasure AI will keep hosting events where we learn, discuss, and share practical knowledge together with the technical community. We look forward to seeing you at the next Treasure AI Night.