Irah is a key contributor to the interface layer of the AI agent stack. By developing protocols like AG-UI and A2UI, the studio addresses how autonomous agents can dynamically generate and manipulate user interfaces based on real-time context. This work is essential for moving beyond the "chatbot" paradigm into fully agentic applications where the software UI adapts to the agent's current task and the user's intent.

Specifically, Irah is an early adopter and implementer of the Model Context Protocol (MCP), using it to bridge the gap between design documentation and model execution. This makes Irah highly relevant to developers building agentic workflows that require sophisticated, production-ready frontend output. The studio champions the concept of Generative UI (GenUI), providing the architectural patterns needed to maintain design integrity while allowing for agent-driven non-determinism.

The interface of the singularity

Irah is the experimental outlet for Hari Periyasamy, an AI Frontend Engineer currently at Shopify and formerly of Amazon Ads. The project functions as a research lab for AI-native frontend architectures, moving beyond the static patterns of traditional web development into what Periyasamy describes as the "near the singularity" era of computing. The core challenge Irah addresses is the inherent friction between deterministic design systems and the non-deterministic outputs of large language models. While most AI applications treat the UI as a simple wrapper for a chat box, Irah treats the interface as a dynamic, agent-orchestrated layer.

Periyasamy brings experience from high-scale environments like Amazon, where he developed AI-powered anomaly detection dashboards and real-time telemetry systems for CloudWatch. This background in production-grade reliability informs the work at Irah, where the focus is on creating "high-integrity" signals. The studio is particularly active in the burgeoning Model Context Protocol (MCP) space, where it has pioneered a contextual UI bridge. This system allows AI models to directly leverage design documentation to generate UI components, effectively closing the loop between a model's intent and the final rendered code.

The Agent-to-UI protocol stack

The most tangible outputs from Irah are three experimental protocols: AG-UI, A2A (Agent-to-Agent), and A2UI (Agent-to-UI). These protocols allow multiple agents to collaborate on analyzing a user's behavior and intent in real-time, subsequently generating a personalized interface. Unlike traditional A/B testing or server-driven UI (SDUI), this approach does not rely on pre-defined templates. Instead, it uses the design system as a set of constraints for the agent, ensuring that the generated output remains brand-compliant and functional while being tailored to specific visitor signals.

This work is founded on the principle that as AI agents become the primary users of software, the nature of the frontend must change. Irah advocates for a Generative UI (GenUI) architecture where the backend provides data and the agent provides the logic for the interface's structure and behavior. This is supported by Periyasamy's previous work at Amazon Ads, where he implemented stream-based data ingestion and backend-for-frontend (BFF) middleware using GraphQL to optimize client-server communication.

Bridging models and production code

The project occupies a niche between research and practical implementation. While many in the agent space focus on the reasoning capabilities of models, Irah focuses on the "last mile" problem of how those models interact with human users and other software systems. The use of Atomic Design principles and Web Components ensures that even agent-generated UIs maintain modularity and performance. By applying observability and telemetry to these non-deterministic interfaces, Irah provides a framework for measuring the effectiveness of agent-driven UX.

Based in the United States and deeply integrated with the Shopify ecosystem, Periyasamy represents a new class of engineer specialized in AI-native frontends. The studio's emphasis on MCP support suggests a commitment to open standards in the agent ecosystem, moving away from closed-loop AI features toward a more interoperable and agent-orchestrated web.